Objective and High-Precision Fabric Wrinkle Assessment using 3D Point Cloud Data and Deep Learning Techniques

Objective and High-Precision Fabric Wrinkle Assessment using 3D Point Cloud Data and Deep Learning Techniques | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Objective and High-Precision Fabric Wrinkle Assessment using 3D Point Cloud Data and Deep Learning Techniques Zhijie Yuan, Zijun Shen, Md All Amin Newton, Binjie Xin, Jing Zhang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5220842/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract This study presents a novel, high-precision method for objectively assessing fabric wrinkles utilizing 3D point cloud data and advanced deep learning techniques. Employing the EinScan-SP scanner, we generated a comprehensive dataset comprising 44 samples from 12 distinct fabric types. The intricate surface topologies of these fabrics were analyzed using PointNet, PointNet++, and PointNet++_MSG architectures, demonstrating their sensitivity to subtle wrinkle details. Our non-contact, automated approach significantly improves upon traditional wrinkle assessment techniques, offering an accurate and reliable means of quality control in the textile industry. The methodology's effectiveness was validated through comparative analysis, showcasing its superiority in terms of accuracy and repeatability. This study establishes a new benchmark for the precise evaluation of fabric surface characteristics, facilitating advancements in textile quality standards. The code and datasets are publicly available at https://github.com/YuanZhijie/FabricPointNet . 3D Point Cloud Data Fabric Surface Wrinkle Quality Control PointNet + + Network Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 1. Introduction Surface wrinkle quality is a critical standard across various industries, including wearable technology[ 1 ], furniture manufacturing[ 2 ], and automotive interior design[ 3 ]. High standards for fabric appearance not only influence the aesthetic appeal of products[ 4 ] but also their functionality and lifespan[ 5 ]. Ensuring accurate and objective wrinkle assessments is essential, as it directly impacts the quality and usability of the final product. Traditional methods, both contact and non-contact, are limited in repeatability and subject to subjective bias, highlighting the need for more reliable and standardized measurement techniques. Moreover, achieving precision in wrinkle characterization is vital, especially in diverse and large-scale production environments, to uphold high-quality standards industry-wide. Currently, the methods for assessing fabric wrinkles primarily fall into two categories: contact and non-contact. For contact measurements, research by Shiloh M [ 6 ] focuses on the evaluation of seam-puckering. The SAWTRI wrinkle meter is used, comparing the lengths of wrinkled and unwrinkled seams to calculate a "wrinkle index" to assess the degree of surface wrinkling. This limitation significantly undermines the applicability and reliability of such methods in a wide range of practical applications. Studies like Galuszynski's [ 7 ], which investigates the effects of fabric structure on needle resistance, also offer insights into fabric durability and quality. Sun[ 8 ] introduced a double extraction method for evaluating fabric wrinkling based on residual force–displacement curves, providing a novel perspective on mechanical characterization. Lu[ 9 ] presented an integrated system to assess textile performance by examining shape retention alongside other fabric properties. Zhou[ 10 ], [ 11 ] utilized advanced machine learning algorithms for classifying and rating wrinkle levels, highlighting the potential for technology-driven assessments in textile quality. Collectively, these studies showcase a significant shift towards more precise and efficient methods for fabric analysis, aiming to enhance both aesthetic and functional qualities of textiles. Despite these advancements, a notable drawback of these methods is their poor repeatability, with results often influenced by the operator's subjective assessment [ 12 ], [ 13 ]. In the field of non-contact measurements, early researchers utilized photoelectric devices to project light onto the wrinkled fabric surface, forming a shadow pattern. By measuring this projection pattern, they obtained the ratio of the surface wrinkle curve length to the seam length, which served as an objective evaluation index to determine the fabric's level of flatness[ 14 ]. In recent years, scholars have used different configurations of lighting and image capture equipment for data collection. Chen Lili [ 15 ] used a camera to collect seam samples at a fixed shooting distance. Meanwhile, Mohri M and others[ 16 ] employed the Radon Transform for the objective evaluation of wrinkled fabric, and Yang X.B and Huang X.B[ 17 ] proposed a photometric stereo method to further analyze fabric wrinkles. These technological advancements have greatly enhanced the precision and reliability of fabric wrinkle analysis, providing valuable tools for quality assessment in textile manufacturing. With the rapid development of deep learning technology [ 18 ]–[ 20 ] and the widespread use of point cloud data [ 21 ], deep learning point cloud evaluation networks have emerged as a powerful tool for fabric wrinkle analysis. Recent advances have focused on developing objective, efficient models using deep learning and optimization algorithms. Shen et al. [ 22 ] proposed a Point Cloud Upsampling Generative Adversarial Network which can improve the granularity of wrinkle characterization, addressing the limitations of conventional methods. Zhou et al. [ 23 ] introduced an artificial hummingbird algorithm with DarkNet19 for enhanced feature abstraction, significantly improving classification accuracy. Liu et al. [ 24 ] developed S2ANet, integrating local spectral and spatial point grouping to refine point cloud processing. These innovations illustrate a shift towards more accurate, objective assessments of fabric quality, highlighting ongoing efforts to enhance textile industry standards. To address the limitations of existing methods and to further improve the measurement and evaluation accuracy of fabric surface characteristics, this study proposes a novel approach by integrating high-precision 3D point cloud data with deep learning algorithms. At the core of this method is the use of an EinScan-SP 3D scanner, which captures detailed and high-precision point cloud data, accurately representing the fabric's surface morphology. This data is then processed using advanced point cloud evaluation networks, designed to overcome the shortcomings of traditional methods, such as low processing efficiency, poor robustness to noise, and subjective bias. By employing this innovative approach, the study offers a more accurate and objective analysis of fabric surface features, setting a new standard in textile quality control. To validate the effectiveness of the proposed methodology, a comparative analysis with previously used methods was conducted, as shown in Table 6. The results demonstrate that the methodology used in this study generally surpasses existing methods in terms of accuracy and reliability. 2. Methodology 2.1. Sample preparation Forty-four fabric samples were prepared and classified, undergoing detailed analysis for fabric smoothness using the EinScan-SP 3D scanner. These samples were categorized into 12 different types based on their surface texture characteristics. The preparation of these fabrics adhered strictly to the standardized procedures necessary for high-resolution 3D scanning. This included conditioning the samples in a controlled environment to ensure consistency and repeatability in the data collection process. The focus of this assessment was to capture detailed 3D images of the fabric surfaces, which are essential for accurate analysis of wrinkle characteristics. 2.2. Digital System for Point Cloud Collection The point cloud acquisition system used in this study primarily consists of the EinScan-SP 3D scanner and a high-speed GPU (NVIDIA GTX 3080). The EinScan-SP offers extremely high scanning precision, with single-shot accuracy reaching up to ≤ 0.05 mm, thus enabling the capture of minute details of the fabric surface. The device has a maximum scanning volume of 200 mm × 200 mm × 200 mm and a single scan time of less than 45 seconds, implying that a substantial amount of data can be collected in a relatively short period. The specific parameters are presented in Table 1. The high resolution and accuracy of this scanning system are crucial for effectively capturing the complex geometries and fine textures of fabric surfaces, which are essential for an in-depth analysis of wrinkle characteristics. Table.1 EinScan-SP V2 1 instrument paprameters 2 Indicator Parameters Single Accuracy ≤ 0.05 mm Minimum Scanning Volume 30 mm × 30 mm × 30 mm Maximum Scanning Volume 200 mm × 200 mm × 200 mm Single Scan Range 200 × 150 mm Scan Speed < 45 sec Point Distance 0.17 mm ~ 0.2 mm [1] EinScan-SP V2 product information: https://www.einscan.com/einscan-sp/ (accessed on 12 June 2024). [2] Detailed parameters of the equipment: https://files.shining3d.com/hubfs/Brochures/EinScan/EinScan%20SE%20SP/%E3%80%90EN%E3%80%91EinScan%20SE-SP%20V2%20-%20V0.2%2020220718.pdf (accessed on 12 June 2024). The integration of the EinScan-SP scanner and the NVIDIA GTX 3080 graphics card allows for rapid and efficient data processing, ensuring that the high volume of point cloud data collected can be handled smoothly. The acquisition of point clouds is performed in a dark room to eliminate interference from external light sources, which could affect the scanning results. The scanning setup and its configuration are critical for obtaining high-quality data, which is essential for precise and detailed analysis of the fabric's surface features. 2.3. Dataset generation method 2.3.1. Data acquisition and Pre-Processing The EinScan-SP scanner is capable of capturing both 3D and color information, which can be exported in various formats optimized for detailed analysis. For this experiment, data in point cloud format were selected to maximize the accuracy and resolution of the surface details captured. This format is particularly effective in retaining the intricate textures and geometries of the fabric surfaces which are crucial for analyzing wrinkle characteristics. The pre-processing steps included cleaning the data of noise and irrelevant features to ensure that the scans provided a clear and accurate representation of the fabric’s intrinsic properties. The point cloud data obtained, as depicted in Fig. 3 , encompasses a wealth of information, including the shape, texture, and color of the fabric surface. This data is unstructured and may contain noise and missing data points. Consequently, a series of preprocessing steps were performed to refine and stabilize the point cloud data. These steps are crucial for ensuring that the detailed and precise measurements needed for an effective analysis of the fabric's wrinkle characteristics are reliable and accurate. After obtaining the point cloud, it was essential to refine the data to ensure optimal clarity and detail for fabric analysis. This refinement process involved aligning and merging multiple scans to create a comprehensive view of each fabric sample. Accurate alignment is crucial as it significantly affects the quality of the 3D model and thereby the reliability of the subsequent analysis of fabric wrinkles. The final point cloud represents a detailed and precise digital model of the fabric surface, enabling in-depth study of the intricate patterns and textures that characterize fabric wrinkles. These processes are shown in Fig. 3 . 2.3.2. Point cloud downsampling By performing the aforementioned operations, 3D scanning devices can capture data comprising millions of points, as shown in Table 2. This enables the detailed capture of surface features on fabrics, essential for accurately characterizing wrinkle patterns and textures. Given that fabric wrinkles generally have a scale of 0.5 mm or more, the actual acquisition accuracy of 0.17 mm − 0.2 mm implies that the device captures excessive data points, which are then meticulously filtered to focus solely on relevant wrinkle characteristics. Storing, processing, and displaying this data efficiently is crucial for maintaining the integrity and usability of the analytical results, ensuring that the subtle nuances of fabric wrinkles are accurately represented. In Table 2, data acquired for different fabric types using the 3D scanning device is presented, listing 12 various fabrics including linen, cotton, and silk, with respective counts of vertices and faces. Although the data magnitude is somewhat similar across different fabrics, the unique composition and texture characteristics of each fabric type subtly influence the overall processing and analysis of wrinkle detection. For instance, the structured nature of linen may result in different wrinkle characteristics compared to cotton or silk, affecting how point cloud data is handled. Table.2 Different type fabric data acquired by the 3D scanning device No. Type Composition Number of Vertices Number of Faces 1 Floral leaf linen Linen 1003469 2002010 2 Blue floral Cotton 878498 1752552 3 Linen cat Linen 1007421 2009872 4 Star Cotton 972984 1941391 5 Red checkered Cotton 984535 1964549 6 Green leaf Cotton 810187 1616273 7 Leaf linen Linen 979521 1954050 8 Ink-style Silk 981711 1959410 9 Yellow leaf Cotton 1021183 2037932 10 Striped Cotton 941923 1878879 11 Wave Cotton 1121753 2239170 12 Elephant patterned Silk 636563 1270832 Figure 4 illustrates the WR-1 sample under different viewing angles, each offering unique insights into the fabric's surface features. Normal shooting (Fig. 4 a) provides a baseline view of the fabric, showing the general appearance and texture. Camera magnification (Fig. 4 b) zooms into the fabric, offering a closer look at the surface details and wrinkle characteristics. Imaging microscopy (Fig. 4 c) provides an even more detailed examination of the fabric's surface, revealing intricate textures and fine wrinkles that are not visible in the normal shooting view. Finally, the point clouds after 3D reconstruction (Fig. 4 d) present a comprehensive and precise 3D model of the fabric's surface, capturing every nuance of the wrinkle and texture. These different angles and techniques collectively contribute to a fuller understanding of the fabric's surface characteristics, aiding in the accurate evaluation of fabric wrinkles. Farthest Point Sampling (FPS) is a prevalent strategy in point cloud processing, particularly when there is a need to reduce data or perform down-sampling. The core concept behind FPS is to select points that are farthest from the already sampled points as new sampling points, thereby ensuring a uniform distribution of sampling points across the entire point cloud. This technique is crucial for maintaining the quality and integrity of the data, especially when analyzing complex fabric surfaces where precision in detail is essential. The reduced data set still retains the critical information necessary for a detailed examination of the fabric's wrinkle characteristics. In implementing FPS, the process begins by randomly selecting an initial point as the first sampling point, accompanied by the creation of a distance array with elements set to infinity. This array is crucial as it stores the distance from each point to the nearest point in the already sampled set. Sampling then proceeds through iterative steps. In each iteration, distances from the unsampled points to the nearest sampled point are recalculated, and the point with the maximum distance is chosen as the new sampling point. This systematic approach ensures that the sampled points optimally cover the entire fabric surface, which is essential for a detailed and comprehensive analysis of the fabric's wrinkle characteristics. The sampling method can be mathematically represented by the following equation: $$\:{D}_{i}=min({D}_{i},||{P}_{i}-{P}^{{\prime\:}}|{|}^{2})$$ 1 where D i is the distance from the i-th point to the nearest point in the already sampled point set, P i is the coordinates of the i-th point, and P' is a point from the already sampled point set. The index corresponding to argmax(D i ) is chosen as the next sampling point. This FPS strategy yields a downsampled point cloud along with the corresponding indices of the sampled points. In scenarios involving large datasets or limited computational resources, this sampling method proves instrumental in reducing the data volume while still capturing the essential structure and details of the point cloud. Such efficiency is crucial in point cloud processing for fabric analysis. The downsampled point clouds effectively preserve the structural integrity and detailed textural information of the fabric surfaces, essential for accurately characterizing wrinkle patterns, as illustrated in Fig. 6 . 2.3.3. Point cloud normalization Before processing point cloud data, certain preprocessing steps are necessary, among which normalization is a key component. Normalization ensures that data of different scales and units can be correctly interpreted and compared. Firstly, the point cloud data are normalized by calculating the mean and standard deviation using the following formulas, where \(\:{p}_{i}\) represents each point and N is the total number of points: \(\:mean=\frac{\sum\:{p}_{i}}{N}\) (2) \(\:std=\sqrt{\frac{\sum\:({p}_{i}-mean{)}^{2}}{N}}\) (3) Then, the point cloud data are processed using zero-mean normalization (also known as standardization) to ensure the normalized data has a mean of 0 and a standard deviation of 1. This standardization is crucial as it facilitates a uniform platform for analyzing intricate wrinkle characteristics by mitigating scale disparities across different scans. The formula is as follows: $$\:points\_ormalized=\frac{points-mean}{std}$$ 4 Next, the normals of the point cloud are processed. The maximum Euclidean norm of all the normals is calculated using the following formula, where \(\:{p}_{i}\) is each normal vector: $$\:max\_norm=max\left(\sqrt{\sum\:{p}_{i}^{2}}\right)$$ 5 To achieve consistent analysis quality across all normals, they are normalized using this maximum norm by dividing each normal vector by the maximum norm: $$\:normals\_normalized=\frac{normals}{max\_norm}$$ 6 Finally, the normalized point cloud data and normals are recombined to form a complete point cloud dataset. The normalization process ensures that the scale of the point cloud data and normal values is consistent, which is crucial for subsequent analysis and processing aimed at wrinkle characterization. As illustrated in Fig. 6 , blue indicates points closest to the sample, and yellow marks those farthest from the reference plane of the sample. Due to changes in the reference plane caused by downsampling and normalization—particularly when downsampling reduces data to a tenth of its original scale—the appearance of the point cloud might alter when viewed from a fixed angle. Nonetheless, it is evident that the essential characteristics of the fabric wrinkles are preserved, allowing for accurate analysis of wrinkle patterns and textures after these processing steps. 2.3.4. Data Augmentation Due to the limited scale of the original dataset, deep learning models may struggle to learn the underlying patterns and complexities of fabric wrinkles, thus affecting their ability to generalize these features[ 25 ]. Small-scale datasets can lead to overfitting and poor performance on unseen fabric data[ 26 ], while imbalanced datasets may further impair the models' ability to recognize certain types of wrinkles due to insufficient data [ 27 ]. To address these issues, four methods were employed to enhance the dataset's robustness in representing diverse wrinkle characteristics. A) Rotation : Rotation is a crucial operation in point cloud data processing. It involves converting the rotation angle from degrees to radians and performing a rotation using a 3x3 rotation matrix. In this study, 8 rotations were performed on the fabric point cloud using a 45-degree rotation angle to ensure comprehensive analysis from multiple perspectives, which is vital for understanding the varied wrinkle patterns on the fabric surface. B) Translation Translation is a geometric transformation that allows for the movement of the entire point cloud along specified directions without changing their shape or orientation. This operation is important for aligning and comparing different sections of fabric when analyzing wrinkle characteristics. C) Noise Adding random noise to the data helps assess the robustness of the point cloud processing approach specifically for fabric wrinkle analysis. Gaussian noise, with the same shape as the point cloud, is generated and added to the original points. The surface normals are unaffected by the noise to maintain the original surface orientation. In this study, five Gaussian noises with random frequencies were generated and two of them were randomly applied to the point cloud for each iteration, as shown in Fig. 7 (d). This operation is crucial for ensuring the fabric's wrinkle characteristics can be reliably analyzed under various noise conditions. D) Scaling Scaling allows for changing the size of objects or adjusting the distance between objects and observers, which is essential for analyzing how these variations affect the visibility and characteristics of fabric wrinkles. Each point in the point cloud is multiplied by a scaling factor, while the surface normals remain unchanged to maintain accurate orientation of the fabric surfaces. In this study, five scaling factors were selected: 1.2, 1.1, 1.0, 0.9, and 0.8, and two of them were randomly applied to the point cloud for each iteration, as demonstrated in Fig. 7 (e). This operation is crucial for examining the effects of scale on fabric wrinkles, ensuring that the analysis captures the true nature of wrinkle patterns under different conditions. 2.4. Evaluation models Evaluation neural networks for processing fabric wrinkle point cloud data possess several key characteristics: A) Handling unordered data : Point cloud data is unordered, allowing for arbitrary point arrangements without affecting the overall shape information. Neural networks handle this through pooling operations using symmetric functions such as maximum or average value[ 28 ]–[ 30 ]. B) Invariance to rotation and scaling : Point cloud data can vary in rotation and scaling. Neural networks achieve invariance to transformations by incorporating alignment or normalization steps within the network [ 31 ]–[ 33 ]. C) Integration of Local and Global Features : Point cloud data contains both local and global shape information, as well as multi-scale feature learning [ 34 ]. D) Efficient Spatial Search and Computation : Point cloud data consists of a large number of points, requiring efficient spatial search and computation. This is achieved through optimization algorithms and data structures such as K-D trees and octrees[ 35 ]–[ 38 ]. Neural networks for point cloud data processing excel in various applications, including 3D object recognition, semantic segmentation, and point cloud generation. The PointNet series of networks handle point cloud data directly by utilizing symmetric functions for global feature learning and employing feature transformation networks for alignment and normalization. This approach ensures invariance to point cloud arrangement, rotation, and translation, while also adapting well to variations in point cloud size and density. In this study, the PointNet series of networks is used to evaluate the surface smoothness of textile materials after preprocessing the point cloud data. 2.4.1. PointNet architecture PointNet[ 39 ] is a neural network designed to directly process point cloud data. Its primary characteristic lies in its ability to extract useful features directly from raw point cloud data, thereby eliminating the need for complex data preprocessing or transformation. The network structure of PointNet primarily consists of two parts: the Input Transform and the Evaluation Network, as shown in Fig. 9 . A) Input Transform : This is a miniaturized version of PointNet, with the objective of learning an alignment matrix to align the input point cloud to a normalized coordinate space. This process ensures the network's invariance to the rotation and translation of the input point cloud. B) Evaluation Network : The Evaluation Network forms the main body of PointNet. It initially extracts the features of each point through a series of Multi-Layer Perceptrons (MLP), then obtains global features through a max pooling layer. The max pooling layer ensures the network's invariance to the order of the input points. Finally, the global features are processed through another series of MLPs to obtain the final evaluation result. During the implementation process, PointNet incorporates a T-Net at each stage of feature extraction. This enhances the network's invariance to the rotation and scaling of the input point cloud. 2.4.2. PointNet + + architecture PointNet + + is a network that builds upon the foundation of PointNet to address some of the challenges encountered when dealing with complex structures and large-scale point cloud data, as shown in Fig. 10 . The main improvements in PointNet + + include the introduction of hierarchical feature learning and local structure modeling of point clouds. A) Hierarchical Point Set Feature Learning : Instead of directly performing feature learning on the entire point cloud, PointNet + + adopts a different approach. It first divides the point cloud into multiple smaller regions and applies PointNet for feature learning on each of these regions individually. This is achieved by sampling and aggregating the neighborhood of each point. The hierarchical feature learning method enables PointNet + + to better handle large-scale and complex point cloud data. B) Evaluation Network : When computing local features, PointNet + + considers not only the characteristics of each point itself but also the characteristics of its neighboring points. The local features obtained from these regions are then processed using Multi-Layer Perceptrons (MLPs) to derive global features. This approach captures the local structural information of the point cloud, thereby improving the performance of the model. Compared to PointNet, PointNet + + exhibits superior performance in handling complex structures and large-scale point cloud data. However, it also comes with higher computational complexity and storage requirements. In general, PointNet + + typically performs a single random farthest point sampling and sets a grouping radius to extract PointNet features from the grouped point clouds. However, the Multi-Scale Grouping (MSG) approach utilizes a different strategy, as shown in Fig. 11 . It also conducts a random farthest point sampling, but with multiple grouping radii specified for each group. Subsequently, PointNet features are extracted individually for each group. Finally, the point cloud features extracted at different grouping radii are concatenated to obtain the final features output by the Set Abstraction (SA) module. 3. Results and discussion 3.1. Experimental environment The configuration for the network training and testing environment is as follows: Processor: 12th Gen Intel(R) Core(TM) i7–12700K; Memory: 32 GB; Operating system: Windows 11 64-bit; GPU: NVIDIA GeForce RTX 3080; Programming platform: Python 3.9; Deep learning framework: PyTorch 1.6.0. 3.2. Dataset Prepared In the experiment, the point cloud dataset for each type of fabric surface wrinkle is divided into a training set and a test set at a 4:1 ratio, as shown in Table 3. A total of 44 original point clouds of 12 types of fabric surface wrinkle were collected, as shown in Fig. 12 . Each original point cloud is augmented to generate 40 augmented point clouds, resulting in a total of 1760 samples, as shown in Fig. 13 . From each group of 40 augmented point clouds, 32 point clouds are randomly selected, resulting in a total of 1408 point clouds used as training samples. The remaining 8 augmented point clouds from each original point cloud are used as test samples, resulting in a total of 352 test samples. Table.3 Fabric Surface Wrinkle point cloud dataset Type Number of training Number of testing WR-1 128 32 WR-2 256 64 WR-3 320 80 WR-4 448 112 WR-5 256 64 3.3. Evaluation results and analysis 3.3.1. PointNet Figure 14 illustrates the training progress of the PointNet transfer learning model with different configurations: (a) PointNet (without normal) and (b) PointNet. From the experimental results in Fig. 14 a, it is observed that for PointNet (without normal), the accuracy measures start to ascend notably after 110 epochs. However, even reaching up to 200 epochs, the model does not achieve a stable or high accuracy level, failing to surpass 80%. This highlights the initial slow and unstable learning process when normal vectors are not included, reflecting PointNet's challenge in capturing the nuanced features of fabric wrinkles without sufficient directional information provided by normals. In contrast, Fig. 14 b depicts the training progress for the standard PointNet model. Here, the model demonstrates a more robust performance. While the early epochs exhibit some variability, the accuracy quickly approaches optimal levels, demonstrating the effectiveness of including normal vectors in capturing fabric wrinkles more accurately and efficiently. The graph shows a more rapid and stable convergence compared to the variant without normal vectors, indicating that the normal vector inclusion significantly aids in the model's ability to discern detailed surface features and improves overall learning effectiveness. These observations from Fig. 14 are supported by the accuracy data in Table 4 and Table 5, where PointNet with normal vectors consistently shows superior performance in training and evaluation tasks compared to its counterpart without normal vectors. 3.3.2. PointNet++ Figure 15 showcases the training progress of the PointNet + + transfer learning model with comparisons between configurations with and without normal vectors. In Fig. 15 a, when normal vectors are included, PointNet + + demonstrates a rapid learning curve, achieving over 95% training accuracy within just 50 epochs. This is a significant improvement over the standard PointNet model, which reaches similar levels of training accuracy only after approximately 100 epochs as depicted in Fig. 14 b. The enhanced training efficiency of PointNet + + is attributed to its ability to more effectively identify fabric wrinkle areas by incorporating additional components focusing on local features. However, with the inclusion of normal vectors, PointNet + + may experience fluctuations in accuracy during the training process, particularly in test accuracy and class accuracy, which oscillate between 80%-95% as iterations proceed. These fluctuations are likely due to the model's heightened sensitivity to the subtle differences between wrinkled and non-wrinkled areas in the presence of normal vectors. The data shows that while PointNet + + can quickly learn and distinguish different wrinkle grades, the finer details captured by normal vectors introduce variability during the early training phases. In contrast, Fig. 15 b illustrates the training progress of PointNet + + without normal vectors. Notably, even while retaining a higher level of training efficiency, the model demonstrates reduced fluctuations in test accuracy and class accuracy, with these metrics becoming more stable after about 125 epochs. The removal of normal vectors seems to mitigate the model's sensitivity to minute surface variations, leading to a smoother and more predictable learning process while still maintaining a high level of accuracy. 3.3.3. PointNet++_MSG Figure 16 depicts the training progress of the PointNet++_MSG transfer learning model with and without normal vectors, demonstrating its sensitivity to local features and subsequent impact on training stability and accuracy. Figure 16 a, which illustrates the PointNet++_MSG model with normal vectors, shows more pronounced fluctuations in accuracy, oscillating between 50%-90% within 200 epochs, indicating a less stable training process compared to the other models. These fluctuations are attributed to the model's high sensitivity to local information provided by normal vectors, leading to significant changes in accuracy with each iteration, especially when the dataset is complex or contains detailed wrinkle features. In contrast, Fig. 16 b shows the training progress of the PointNet++_MSG model after the removal of normal vectors. Here, the model demonstrates delayed training accuracy, reaching around 92% by the 25th epoch, compared to the 95% achieved by PointNet + + at the same epoch count. However, after 50 epochs, PointNet++_MSG (without normal) exhibits a trend towards stability in both test accuracy and class accuracy, achieving a more stable and reliable performance sooner than PointNet++, which only begins to stabilize after about 125 epochs. This indicates that while the initial learning may be slightly slower without normal vectors, the overall training process becomes more stable and efficient, reducing the wide accuracy fluctuations seen with normal vectors. Table.4 Training and Testing Accuracy of PointNet Series Networks Item Train Accuracy Test Accuracy Best Accuracy PointNet 0.9688 - - PointNet(without normal) 0.5355 0.6818 0.7121 PointNet++ 0.9950 0.7045 0.6750 PointNet++(without normal) 0.9950 - - PointNet++_MSG 0.9950 0.9290 0.9245 PointNet++_MSG(without normal) 0.9979 Table.5 Evaluation Accuracy of PointNet Series Networks Item Evaluate Accuracy BEA (Best Evaluate Accuracy) PointNet(without normal) 0.6989 0.7339 PointNet++ 0.9773 0.9816 PointNet++_MSG 0.9943 0.9938 In order to confirm the validity of the methodology proposed in this paper, a comparative analysis was carried out with the methodology used in previous studies, which is shown in Table 6. The results show that, in general, the methodology used in this study is superior in terms of accuracy and reliability. Table.6 Comparison of evaluation accuracy of different methods Model Train Accuracy Test Accuracy Best Accuracy Evaluate Accuracy BEA SVM[ 18 ] 0.3984 0.4062 0.4062 0.4062 0.4062 Optimized SVM[ 19 ] 0.7195 0.7131 0.7195 0.7131 0.7131 KNN[ 20 ] 0.9247 0.8920 0.9247 0.8920 0.8920 PointNet(without normal) 0.5355 0.6818 0.7121 0.6989 0.7339 PointNet++ 0.9950 0.7045 0.6750 0.9773 0.9816 PointNet++_MSG 0.9950 0.9290 0.9245 0.9943 0.9938 4. Conclusions Utilizing advanced 3D point cloud data and machine learning algorithms, this research achieves precise evaluation of fabric surface wrinkles. A robust dataset was assembled from 44 three-dimensional point cloud samples spanning 12 fabric types and 5 AATCC-WR grades with the aid of the EinScan-SP scanner, enhanced through techniques such as rotation, translation, noise addition, and scaling. Evaluations employing PointNet, PointNet++, and PointNet++_MSG architectures demonstrate that these methods, especially when augmented by normal vectors, significantly outperform traditional wrinkle assessment techniques. The use of optimized multi-scale grouping in PointNet++_MSG effectively minimizes interference from point normals, leading to accurate assessments. These advancements underscore the potential of machine learning in transforming textile quality control by setting new standards for precision and reliability in fabric assessment. The non-contact nature and reduced sensitivity to fabric variations further enhance its applicability across the textile industry, marking a significant step forward in integrating computational technology into traditional practices. Declarations Author Contribution Yuan completed the primary research and drafted the main content. Xin formulated the research direction. Newton reviewed the manuscript for English grammar. Yuan and Shen created the dataset. Zhang refined the research content. All authors reviewed the manuscript. Acknowledgement This project was funded by the National Natural Science Foundation of China (Grant No. 11702169). Data Availability The data supporting the findings of this study are available upon reasonable request from the authors. Code and dataset links in this article: https://github.com/YuanZhijie/FabricPointNet. References M. S. Brown, B. Ashley, and A. Koh, “Wearable technology for chronic wound monitoring: current dressings, advancements, and future prospects,” Frontiers in bioengineering and biotechnology, vol. 6, p. 47, 2018, doi: 10.3389/fbioe.2018.00047 . A. Gurarda, “Seam performance of garments,” in Textile Manufacturing Processes, IntechOpen, 2019. N. Shen, A. Samanta, H. Ding, and W. W. Cai, “Simulating microstructure evolution of battery tabs during ultrasonic welding,” Journal of Manufacturing Processes, vol. 23, pp. 306–314, 2016. R. Narain, A. Samii, and J. F. O’brien, “Adaptive anisotropic remeshing for cloth simulation,” ACM transactions on graphics (TOG), vol. 31, no. 6, pp. 1–10, 2012. X. Chen, M. Yu, and Y. Song, “Optimized Seam-Driven Image Stitching Method Based on Scene Depth Information,” Electronics, vol. 11, no. 12, p. 1876, 2022, doi: 10.3390/electronics11121876 . M. Shiloh, “The evaluation of seam-puckering,” 1971, doi: 10.1080/00405007108630042. S. Galuszynski, “Effect of fabric structure on fabric resistance to needle piercing,” 1986, doi: 10.1177/004051758605600511 . F. Sun, M. Guo, X. Hu, L. Wang, and W. Gao, “Analysis of curve parameters to characterize multidirectional fabric wrinkling by a double extraction method,” Textile Research Journal, vol. 89, no. 15, pp. 2973–2982, Aug. 2019, doi: 10.1177/0040517518805372 . Y. Lu, X. Hu, F. Sun, F. Peng, and W. Gao, “Determination of optimal system parameters to characterize the wrinkle recovery of fabrics by an integrated shape retention evaluation system,” Textile Research Journal, vol. 90, no. 1, pp. 91–100, Jan. 2020, doi: 10.1177/0040517519858770 . Z. Zhou, R. Zhang, J. Zhang, Y. Wang, Z. Zhu, and C. Liu, “Fabric wrinkle level classification via online sequential extreme learning machine based on improved sine cosine algorithm,” Textile Research Journal, vol. 90, no. 17–18, pp. 2007–2021, Sep. 2020, doi: 10.1177/0040517520908072 . Z. Zhou, Z. Ma, Y. Wang, and Z. Zhu, “Fabric wrinkle rating model based on ResNet18 and optimized random vector functional-link network,” Textile Research Journal, vol. 93, no. 1–2, pp. 172–193, Jan. 2023, doi: 10.1177/00405175221117614 . R. Furferi, L. Governi, and Y. Volpe, “Machine Vision-Based Pilling Assessment: A Review,” Journal of Engineered Fibers and Fabrics, vol. 10, no. 3, p. 155892501501000, Sep. 2015, doi: 10.1177/155892501501000320 . P. Yildirim, D. Birant, and T. Alpyildiz, “Data mining and machine learning in textile industry,” WIREs Data Min & Knowl, vol. 8, no. 1, p. e1228, Jan. 2018, doi: 10.1002/widm.1228 . A. M. Bertoldi and D. L. Munden, “The effects of sewing variables on fabric pucker,” Cloth. Res. J, vol. 2, no. 2, p. 81, 1974. L. L. Chen, “Objective evaluation method for sewing flatness of worsted wool fabrics,” J Text Res, vol. 3, pp. 120–125, 2018. M. Mohri, S. H. Ravandi, and M. Youssefi, “Objective evaluation of wrinkled fabric using radon transform,” Journal of the Textile Institute, vol. 96, no. 6, pp. 365–370, 2005, doi: 10.1533/joti.2004.0066 . X. B. Yang and X. B. Huang, “Evaluating fabric wrinkle degree with a photometric stereo method,” Textile Research Journal, vol. 73, no. 5, pp. 451–454, 2003, doi: 10.1177/004051750307300513 . Jingjing Sun, Ming Yao, Bugao Xu, and P. Bel, “Fabric wrinkle characterization and classification using modified wavelet coefficients and support-vector-machine classifiers,” Textile Research Journal, vol. 81, no. 9, pp. 902–913, Jun. 2011, doi: 10.1177/0040517510391702 . N. Deng, Y. Wang, B. Xin, and W. Wang, “A novel objective wrinkle evaluation method for printed fabrics based on multi-view stereo algorithm,” The Journal of The Textile Institute, vol. 113, no. 3, pp. 367–377, Mar. 2022, doi: 10.1080/00405000.2021.1882071 . Y. Wang, N. Deng, and B. Xin, “Investigation of 3D surface profile reconstruction technology for automatic evaluation of fabric smoothness appearance,” Measurement, vol. 166, p. 108264, Dec. 2020, doi: 10.1016/j.measurement.2020.108264 . B. Yang, N. Haala, and Z. Dong, “Progress and perspectives of point cloud intelligence,” Geo-spatial Information Science, pp. 1–17, 2023, doi: 10.1080/10095020.2023.2175478 . B. Shen, L. Li, X. Hu, S. Guo, J. Huang, and Z. Liang, “Point cloud upsampling generative adversarial network based on residual multi-scale off-set attention,” Virtual Reality & Intelligent Hardware, vol. 5, no. 1, pp. 81–91, 2023, doi: 10.1016/j.vrih.2022.08.016 Z. Zhou, Y. Hu, Z. Zhu, and Y. Wang, “Fabric Wrinkle Objective Evaluation Model with Random Vector Function Link Based on Optimized Artificial Hummingbird Algorithm,” J. Nat. Fibers, vol. 20, no. 1, p. 2163026, Dec. 2023, doi: 10.1080/15440478.2022.2163026 . Y. Liu, X. Sun, W. Shao, and Y. Yuan, “S2ANet: Combining local spectral and spatial point grouping for point cloud processing,” Virtual Reality & Intelligent Hardware, vol. 6, no. 4, pp. 267–279, 2024, doi: 10.1016/j.vrih.2023.06.005 . M. Oleynik, A. Kugic, Z. Kasáč, and M. Kreuzthaler, “Evaluating shallow and deep learning strategies for the 2018 n2c2 shared task on clinical text classification,” Journal of the American Medical Informatics Association, vol. 26, no. 11, pp. 1247–1254, 2019, doi: 10.1093/jamia/ocz149 . S. Harrer, P. Shah, B. Antony, and J. Hu, “Artificial intelligence for clinical trial design,” Trends in pharmacological sciences, vol. 40, no. 8, pp. 577–591, 2019, doi: 10.1016/j.tips.2019.05.005 . X. Zhang, J. Chen, Y. Zhou, L. Han, and J. Lin, “A multiple-layer representation learning model for network-based attack detection,” IEEE Access, vol. 7, pp. 91992–92008, 2019, doi: 10.1109/ACCESS.2019.2927465 . W. Zhang, S. Su, B. Wang, Q. Hong, and L. Sun, “Local k-NNs pattern in Omni-Direction graph convolution neural network for 3D point clouds,” Neurocomputing, vol. 413, pp. 487–498, 2020, doi: 10.1016/j.neucom.2020.06.095 . M. D. Phung, C. H. Quach, D. T. Chu, N. Q. Nguyen, T. H. Dinh, and Q. P. Ha, “Automatic interpretation of unordered point cloud data for uav navigation in construction,” in 2016 14th International Conference on Control, Automation, Robotics and Vision (ICARCV), IEEE, 2016, pp. 1–6. doi: 10.1109/ICARCV.2016.7838683 . W. Liu, J. Sun, W. Li, T. Hu, and P. Wang, “Deep learning on point clouds and its application: A survey,” Sensors, vol. 19, no. 19, p. 4188, 2019, doi: 10.3390/s19194188 . K. Al-Manasir and C. S. Fraser, “Registration of terrestrial laser scanner data using imagery,” The Photogrammetric Record, vol. 21, no. 115, pp. 255–268, 2006, doi: 10.1111/j.1477-9730.2006.00379.x . A. Zaganidis, L. Sun, T. Duckett, and G. Cielniak, “Integrating deep semantic segmentation into 3-d point cloud registration,” IEEE Robotics and Automation Letters, vol. 3, no. 4, pp. 2942–2949, 2018, doi: 10.1109/LRA.2018.2848308 . S. A. Eslami et al., “Neural scene representation and rendering,” Science, vol. 360, no. 6394, pp. 1204–1210, 2018, doi: 10.1126/science.aar6170 . Y. Wang, Y. Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon, “Dynamic graph cnn for learning on point clouds,” Acm Transactions On Graphics (tog), vol. 38, no. 5, pp. 1–12, 2019. F. Wang and Z. Zhao, “A survey of iterative closest point algorithm,” in 2017 Chinese Automation Congress (CAC), IEEE, 2017, pp. 4395–4399. doi: 10.1109/CAC.2017.8243553 . Z. Chen et al., “An artificial bee bare-bone hunger games search for global optimization and high-dimensional feature selection,” Iscience, vol. 26, no. 5, 2023, doi: 10.1016/j.isci.2023.106679 . X. Shi, T. Liu, and X. Han, “Improved Iterative Closest Point (ICP) 3D point cloud registration algorithm based on point cloud filtering and adaptive fireworks for coarse registration,” International Journal of Remote Sensing, vol. 41, no. 8, pp. 3197–3220, 2020, doi: 10.1080/01431161.2019.1701211 . C. Wang, M. Ji, J. Wang, W. Wen, T. Li, and Y. Sun, “An improved DBSCAN method for LiDAR data segmentation with automatic Eps estimation,” Sensors, vol. 19, no. 1, p. 172, 2019, doi: 10.3390/s19010172 . J. Xie, Y. Xu, Z. Zheng, S.-C. Zhu, and Y. N. Wu, “Generative pointnet: Deep energy-based learning on unordered point sets for 3d generation, reconstruction and classification,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 14976–14985. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5220842","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":383880694,"identity":"8416ff43-d192-4b05-9a86-ea393e80d669","order_by":0,"name":"Zhijie Yuan","email":"","orcid":"","institution":"Shanghai University of Engineering Sciences","correspondingAuthor":false,"prefix":"","firstName":"Zhijie","middleName":"","lastName":"Yuan","suffix":""},{"id":383880696,"identity":"58fe2f28-8d4c-418e-ad5b-da8d3ecacaa0","order_by":1,"name":"Zijun Shen","email":"","orcid":"","institution":"Shanghai University of Engineering Sciences","correspondingAuthor":false,"prefix":"","firstName":"Zijun","middleName":"","lastName":"Shen","suffix":""},{"id":383880697,"identity":"54dbc724-5407-49bb-8a4e-548596b29451","order_by":2,"name":"Md All Amin Newton","email":"","orcid":"","institution":"Shanghai University of Engineering Sciences","correspondingAuthor":false,"prefix":"","firstName":"Md","middleName":"All Amin","lastName":"Newton","suffix":""},{"id":383880699,"identity":"5dea7693-84a1-4c8e-8552-80e64a086f35","order_by":3,"name":"Binjie Xin","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAyklEQVRIiWNgGAWjYDACZgY2BokKKIeHeC1nSNLCANTC2EaKFoPjvMceWM6rS+znP8D44G0bg7w5IS2SzXzpBpLbDifOnJHAbDi3jcFwZwMBLfzMPGYSktsOJG64wcAmzdvGkGBwgIAWNrCWOXWJ+88fYP9NlBaILQ3MiRsYEtiYidIi2cxjbiBx7LDxjBuJzZJzzkkYbiCkxeD8GbPHEjV1sv39hw9+eFNmI0/QFhBglgBTjA1AQoII9SC1H4hTNwpGwSgYBSMVAAAAVDgHkRyTqAAAAABJRU5ErkJggg==","orcid":"","institution":"Shanghai University of Engineering Sciences","correspondingAuthor":true,"prefix":"","firstName":"Binjie","middleName":"","lastName":"Xin","suffix":""},{"id":383880700,"identity":"091f9b9f-72a2-4c6c-9326-b6799f764169","order_by":4,"name":"Jing Zhang","email":"","orcid":"","institution":"Shanghai University of Engineering Sciences","correspondingAuthor":false,"prefix":"","firstName":"Jing","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2024-10-08 00:08:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5220842/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5220842/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":72161695,"identity":"63329442-ab3d-4a04-999b-8648e539a95d","added_by":"auto","created_at":"2024-12-23 09:47:36","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":43568,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAATCC Wrinkle Recovery Replicas\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/4fcf71f8e968af90a58e33fb.jpg"},{"id":72161210,"identity":"923682f9-598a-491a-9766-008d60ed4fd8","added_by":"auto","created_at":"2024-12-23 09:39:37","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":203351,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePrinciples of Point Cloud Data Acquisition: (a) Schematic of Structured Light Grating; (b) EinScan-SP Device Scanning Module; (c) Actual Scanning Process Grating Screenshot (Illustrated with AATCC WR-1)\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/d3e6e7f3970dffebc7bd3621.png"},{"id":72161206,"identity":"cc4693c9-1c87-4408-9926-786f6fd2a156","added_by":"auto","created_at":"2024-12-23 09:39:36","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":274264,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFlowchart of data acquisition and preprocessing to obtain original point clouds of fabric wrinkle\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/61aa97ac27965be4faebb936.png"},{"id":72160283,"identity":"96afd18b-cce3-4cde-ae2e-f727ba8c7499","added_by":"auto","created_at":"2024-12-23 09:31:36","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":89094,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eWR-1 sample under different viewing angles (a) normal shooting; (b) camera magnification; (c) imaging microscopy; (d) point clouds after 3D reconstruction\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/79c28bad7d3e06fcde533335.png"},{"id":72161696,"identity":"a2b6233b-725e-409f-99b8-40fcb465408a","added_by":"auto","created_at":"2024-12-23 09:47:36","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":208350,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExample diagram of each wrinkle level: (a)-(c) WR-1; (d)-(f) WR-2; (g)-(i) WR-3; (j)-(l) WR-4; (m)-(o) WR-5\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/411990ef4f9c3ea83acd8407.png"},{"id":72161208,"identity":"2eca4bb8-2bb4-4bc3-9aa9-bfbe23548162","added_by":"auto","created_at":"2024-12-23 09:39:36","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":424073,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSchematic diagram of downsampling and normalization of AATCC-WR sample point clouds\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/af31db247be6c074764655f5.png"},{"id":72160286,"identity":"7d11fcbe-f280-4859-97a9-357b25f84fc9","added_by":"auto","created_at":"2024-12-23 09:31:36","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":83558,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eData Augmentation with Operations: (a) Original Sample; (b) Rotation; (c) Translation; (d) Noise; (e) Scaling.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/c4f4a1909a7485eee00b17e2.png"},{"id":72160290,"identity":"42e8be5a-b0e8-4f0f-bfd0-68549c655069","added_by":"auto","created_at":"2024-12-23 09:31:37","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":267175,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eIllustration of Data Augmentation for AATCC-WR Fabric Sample Images\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/b76ded8065fb40eb1dfa6b53.png"},{"id":72160291,"identity":"9514f1e8-1174-42d4-84b0-92141dcb4122","added_by":"auto","created_at":"2024-12-23 09:31:37","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":152592,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePointNet evaluation network main components\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/2fa973445aa85c3c8ce717fc.png"},{"id":72160289,"identity":"9eca378a-9425-4add-9f19-41e99d50bb9c","added_by":"auto","created_at":"2024-12-23 09:31:37","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":208020,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePointNet++ evaluation network main components\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"10.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/f2626cb8006ab4ec4cfb9e6f.png"},{"id":72160298,"identity":"5b9e1747-ad33-4774-915c-7748ec19394c","added_by":"auto","created_at":"2024-12-23 09:31:37","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":84505,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMulti-scale grouping(MSG)\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"11.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/db33aa472ecddcd8fd02b715.png"},{"id":72161211,"identity":"568080ca-2ae6-490f-bd52-584fc81269ff","added_by":"auto","created_at":"2024-12-23 09:39:37","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":442122,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSample display of fabric surface wrinkle\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"12.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/fc9cfdfe25c16ac668dcbd66.png"},{"id":72160296,"identity":"eb077c33-9cdd-4a4a-822d-2991e800e0d8","added_by":"auto","created_at":"2024-12-23 09:31:37","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":425080,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEnhanced display of fabric sample data\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"13.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/798df057c141f3f819544e83.png"},{"id":72160292,"identity":"93257019-6c53-48c7-b1fc-86cc86598970","added_by":"auto","created_at":"2024-12-23 09:31:37","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":257059,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTraining progress of PointNet transfer learning model\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"14.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/f4707faad1f0be8bc9e47758.png"},{"id":72160294,"identity":"f5175ec4-6083-46d8-b38c-326557083d98","added_by":"auto","created_at":"2024-12-23 09:31:37","extension":"png","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":282730,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTraining progress of PointNet++ transfer learning model\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"15.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/40e42c9879217f84266339f9.png"},{"id":72160295,"identity":"d5306fdb-f5ea-4eb3-8f0f-49267e384e16","added_by":"auto","created_at":"2024-12-23 09:31:37","extension":"png","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":293575,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTraining progress of PointNet++_MSG transfer learning model\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"16.png","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/9f323a6befacfae92270c8b6.png"},{"id":98622212,"identity":"e2f8bd8f-cc3e-4dbd-a5df-980c49d370c1","added_by":"auto","created_at":"2025-12-19 16:49:04","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5308773,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5220842/v1/2ce10660-7376-4366-b87f-95d6ff102633.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Objective and High-Precision Fabric Wrinkle Assessment using 3D Point Cloud Data and Deep Learning Techniques","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eSurface wrinkle quality is a critical standard across various industries, including wearable technology[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], furniture manufacturing[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], and automotive interior design[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. High standards for fabric appearance not only influence the aesthetic appeal of products[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e] but also their functionality and lifespan[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Ensuring accurate and objective wrinkle assessments is essential, as it directly impacts the quality and usability of the final product. Traditional methods, both contact and non-contact, are limited in repeatability and subject to subjective bias, highlighting the need for more reliable and standardized measurement techniques. Moreover, achieving precision in wrinkle characterization is vital, especially in diverse and large-scale production environments, to uphold high-quality standards industry-wide.\u003c/p\u003e \u003cp\u003eCurrently, the methods for assessing fabric wrinkles primarily fall into two categories: contact and non-contact. For contact measurements, research by Shiloh M [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] focuses on the evaluation of seam-puckering. The SAWTRI wrinkle meter is used, comparing the lengths of wrinkled and unwrinkled seams to calculate a \"wrinkle index\" to assess the degree of surface wrinkling. This limitation significantly undermines the applicability and reliability of such methods in a wide range of practical applications. Studies like Galuszynski's [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e], which investigates the effects of fabric structure on needle resistance, also offer insights into fabric durability and quality. Sun[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] introduced a double extraction method for evaluating fabric wrinkling based on residual force\u0026ndash;displacement curves, providing a novel perspective on mechanical characterization. Lu[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] presented an integrated system to assess textile performance by examining shape retention alongside other fabric properties. Zhou[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e] utilized advanced machine learning algorithms for classifying and rating wrinkle levels, highlighting the potential for technology-driven assessments in textile quality. Collectively, these studies showcase a significant shift towards more precise and efficient methods for fabric analysis, aiming to enhance both aesthetic and functional qualities of textiles. Despite these advancements, a notable drawback of these methods is their poor repeatability, with results often influenced by the operator's subjective assessment [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn the field of non-contact measurements, early researchers utilized photoelectric devices to project light onto the wrinkled fabric surface, forming a shadow pattern. By measuring this projection pattern, they obtained the ratio of the surface wrinkle curve length to the seam length, which served as an objective evaluation index to determine the fabric's level of flatness[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. In recent years, scholars have used different configurations of lighting and image capture equipment for data collection. Chen Lili [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] used a camera to collect seam samples at a fixed shooting distance. Meanwhile, Mohri M and others[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] employed the Radon Transform for the objective evaluation of wrinkled fabric, and Yang X.B and Huang X.B[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] proposed a photometric stereo method to further analyze fabric wrinkles. These technological advancements have greatly enhanced the precision and reliability of fabric wrinkle analysis, providing valuable tools for quality assessment in textile manufacturing.\u003c/p\u003e \u003cp\u003eWith the rapid development of deep learning technology [\u003cspan additionalcitationids=\"CR19\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u0026ndash;[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] and the widespread use of point cloud data [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], deep learning point cloud evaluation networks have emerged as a powerful tool for fabric wrinkle analysis. Recent advances have focused on developing objective, efficient models using deep learning and optimization algorithms. Shen et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] proposed a Point Cloud Upsampling Generative Adversarial Network which can improve the granularity of wrinkle characterization, addressing the limitations of conventional methods. Zhou et al. [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] introduced an artificial hummingbird algorithm with DarkNet19 for enhanced feature abstraction, significantly improving classification accuracy. Liu et al. [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e] developed S2ANet, integrating local spectral and spatial point grouping to refine point cloud processing. These innovations illustrate a shift towards more accurate, objective assessments of fabric quality, highlighting ongoing efforts to enhance textile industry standards.\u003c/p\u003e \u003cp\u003eTo address the limitations of existing methods and to further improve the measurement and evaluation accuracy of fabric surface characteristics, this study proposes a novel approach by integrating high-precision 3D point cloud data with deep learning algorithms. At the core of this method is the use of an EinScan-SP 3D scanner, which captures detailed and high-precision point cloud data, accurately representing the fabric's surface morphology. This data is then processed using advanced point cloud evaluation networks, designed to overcome the shortcomings of traditional methods, such as low processing efficiency, poor robustness to noise, and subjective bias. By employing this innovative approach, the study offers a more accurate and objective analysis of fabric surface features, setting a new standard in textile quality control. To validate the effectiveness of the proposed methodology, a comparative analysis with previously used methods was conducted, as shown in Table\u0026nbsp;6. The results demonstrate that the methodology used in this study generally surpasses existing methods in terms of accuracy and reliability.\u003c/p\u003e"},{"header":"2. Methodology","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003e2.1. Sample preparation\u003c/h2\u003e\n \u003cp\u003eForty-four fabric samples were prepared and classified, undergoing detailed analysis for fabric smoothness using the EinScan-SP 3D scanner. These samples were categorized into 12 different types based on their surface texture characteristics. The preparation of these fabrics adhered strictly to the standardized procedures necessary for high-resolution 3D scanning. This included conditioning the samples in a controlled environment to ensure consistency and repeatability in the data collection process. The focus of this assessment was to capture detailed 3D images of the fabric surfaces, which are essential for accurate analysis of wrinkle characteristics.\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003e2.2. Digital System for Point Cloud Collection\u003c/h2\u003e\n \u003cp\u003eThe point cloud acquisition system used in this study primarily consists of the EinScan-SP 3D scanner and a high-speed GPU (NVIDIA GTX 3080). The EinScan-SP offers extremely high scanning precision, with single-shot accuracy reaching up to \u0026le;\u0026thinsp;0.05 mm, thus enabling the capture of minute details of the fabric surface. The device has a maximum scanning volume of 200 mm \u0026times; 200 mm \u0026times; 200 mm and a single scan time of less than 45 seconds, implying that a substantial amount of data can be collected in a relatively short period. The specific parameters are presented in Table\u0026nbsp;1. The high resolution and accuracy of this scanning system are crucial for effectively capturing the complex geometries and fine textures of fabric surfaces, which are essential for an in-depth analysis of wrinkle characteristics.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eTable.1 EinScan-SP V2\u003c/strong\u003e \u003csup\u003e1\u003c/sup\u003e \u003cstrong\u003einstrument paprameters\u003csup\u003e2\u003c/sup\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Taba\" border=\"1\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eIndicator\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eParameters\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSingle Accuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026le;\u0026thinsp;0.05 mm\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMinimum Scanning Volume\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e30 mm \u0026times; 30 mm \u0026times; 30 mm\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMaximum Scanning Volume\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e200 mm \u0026times; 200 mm \u0026times; 200 mm\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSingle Scan Range\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e200 \u0026times; 150 mm\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eScan Speed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;45 sec\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePoint Distance\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.17 mm\u0026thinsp;~\u0026thinsp;0.2 mm\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003cp\u003e[1]\u0026nbsp;EinScan-SP V2 product information: https://www.einscan.com/einscan-sp/ (accessed on 12 June 2024).\u003c/p\u003e\n\u003cp\u003e[2] Detailed parameters of the equipment: \u0026nbsp;https://files.shining3d.com/hubfs/Brochures/EinScan/EinScan%20SE%20SP/%E3%80%90EN%E3%80%91EinScan%20SE-SP%20V2%20-%20V0.2%2020220718.pdf (accessed on 12 June 2024).\u003c/p\u003e\n \u003cp\u003eThe integration of the EinScan-SP scanner and the NVIDIA GTX 3080 graphics card allows for rapid and efficient data processing, ensuring that the high volume of point cloud data collected can be handled smoothly. The acquisition of point clouds is performed in a dark room to eliminate interference from external light sources, which could affect the scanning results. The scanning setup and its configuration are critical for obtaining high-quality data, which is essential for precise and detailed analysis of the fabric\u0026apos;s surface features.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\n \u003ch2\u003e2.3. Dataset generation method\u003c/h2\u003e\n \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e\n \u003ch2\u003e2.3.1. Data acquisition and Pre-Processing\u003c/h2\u003e\n \u003cp\u003eThe EinScan-SP scanner is capable of capturing both 3D and color information, which can be exported in various formats optimized for detailed analysis. For this experiment, data in point cloud format were selected to maximize the accuracy and resolution of the surface details captured. This format is particularly effective in retaining the intricate textures and geometries of the fabric surfaces which are crucial for analyzing wrinkle characteristics. The pre-processing steps included cleaning the data of noise and irrelevant features to ensure that the scans provided a clear and accurate representation of the fabric\u0026rsquo;s intrinsic properties.\u003c/p\u003e\n \u003cp\u003eThe point cloud data obtained, as depicted in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e, encompasses a wealth of information, including the shape, texture, and color of the fabric surface. This data is unstructured and may contain noise and missing data points. Consequently, a series of preprocessing steps were performed to refine and stabilize the point cloud data. These steps are crucial for ensuring that the detailed and precise measurements needed for an effective analysis of the fabric\u0026apos;s wrinkle characteristics are reliable and accurate.\u003c/p\u003e\n \u003cp\u003eAfter obtaining the point cloud, it was essential to refine the data to ensure optimal clarity and detail for fabric analysis. This refinement process involved aligning and merging multiple scans to create a comprehensive view of each fabric sample. Accurate alignment is crucial as it significantly affects the quality of the 3D model and thereby the reliability of the subsequent analysis of fabric wrinkles. The final point cloud represents a detailed and precise digital model of the fabric surface, enabling in-depth study of the intricate patterns and textures that characterize fabric wrinkles. These processes are shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e\n \u003ch2\u003e2.3.2. Point cloud downsampling\u003c/h2\u003e\n \u003cp\u003eBy performing the aforementioned operations, 3D scanning devices can capture data comprising millions of points, as shown in Table\u0026nbsp;2. This enables the detailed capture of surface features on fabrics, essential for accurately characterizing wrinkle patterns and textures. Given that fabric wrinkles generally have a scale of 0.5 mm or more, the actual acquisition accuracy of 0.17 mm \u0026minus;\u0026thinsp;0.2 mm implies that the device captures excessive data points, which are then meticulously filtered to focus solely on relevant wrinkle characteristics. Storing, processing, and displaying this data efficiently is crucial for maintaining the integrity and usability of the analytical results, ensuring that the subtle nuances of fabric wrinkles are accurately represented.\u003c/p\u003e\n \u003cp\u003eIn Table\u0026nbsp;2, data acquired for different fabric types using the 3D scanning device is presented, listing 12 various fabrics including linen, cotton, and silk, with respective counts of vertices and faces. Although the data magnitude is somewhat similar across different fabrics, the unique composition and texture characteristics of each fabric type subtly influence the overall processing and analysis of wrinkle detection. For instance, the structured nature of linen may result in different wrinkle characteristics compared to cotton or silk, affecting how point cloud data is handled.\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eTable.2 Different type fabric data acquired by the 3D scanning device\u003c/strong\u003e\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003cdiv class=\"colspec\" align=\"char\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003ctable id=\"Tabb\" border=\"1\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNo.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eType\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eComposition\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNumber of Vertices\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNumber of Faces\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFloral leaf linen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLinen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1003469\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2002010\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBlue floral\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCotton\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e878498\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1752552\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLinen cat\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLinen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1007421\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2009872\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStar\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCotton\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e972984\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1941391\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRed checkered\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCotton\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e984535\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1964549\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGreen leaf\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCotton\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e810187\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1616273\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLeaf linen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLinen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e979521\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1954050\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eInk-style\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSilk\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e981711\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1959410\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYellow leaf\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCotton\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1021183\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2037932\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStriped\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCotton\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e941923\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1878879\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWave\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCotton\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1121753\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2239170\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eElephant patterned\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSilk\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e636563\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1270832\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e illustrates the WR-1 sample under different viewing angles, each offering unique insights into the fabric\u0026apos;s surface features. Normal shooting (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003ea) provides a baseline view of the fabric, showing the general appearance and texture. Camera magnification (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003eb) zooms into the fabric, offering a closer look at the surface details and wrinkle characteristics. Imaging microscopy (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003ec) provides an even more detailed examination of the fabric\u0026apos;s surface, revealing intricate textures and fine wrinkles that are not visible in the normal shooting view. Finally, the point clouds after 3D reconstruction (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003ed) present a comprehensive and precise 3D model of the fabric\u0026apos;s surface, capturing every nuance of the wrinkle and texture. These different angles and techniques collectively contribute to a fuller understanding of the fabric\u0026apos;s surface characteristics, aiding in the accurate evaluation of fabric wrinkles.\u003c/p\u003e\n \u003cp\u003eFarthest Point Sampling (FPS) is a prevalent strategy in point cloud processing, particularly when there is a need to reduce data or perform down-sampling. The core concept behind FPS is to select points that are farthest from the already sampled points as new sampling points, thereby ensuring a uniform distribution of sampling points across the entire point cloud. This technique is crucial for maintaining the quality and integrity of the data, especially when analyzing complex fabric surfaces where precision in detail is essential. The reduced data set still retains the critical information necessary for a detailed examination of the fabric\u0026apos;s wrinkle characteristics.\u003c/p\u003e\n \u003cp\u003eIn implementing FPS, the process begins by randomly selecting an initial point as the first sampling point, accompanied by the creation of a distance array with elements set to infinity. This array is crucial as it stores the distance from each point to the nearest point in the already sampled set. Sampling then proceeds through iterative steps. In each iteration, distances from the unsampled points to the nearest sampled point are recalculated, and the point with the maximum distance is chosen as the new sampling point. This systematic approach ensures that the sampled points optimally cover the entire fabric surface, which is essential for a detailed and comprehensive analysis of the fabric\u0026apos;s wrinkle characteristics. The sampling method can be mathematically represented by the following equation:\u003c/p\u003e\n \u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\n \u003cdiv id=\"FileID_Equ1\" class=\"mathdisplay\"\u003e$$\\:{D}_{i}=min({D}_{i},||{P}_{i}-{P}^{{\\prime\\:}}|{|}^{2})$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003ewhere D\u003csub\u003ei\u003c/sub\u003e is the distance from the i-th point to the nearest point in the already sampled point set, P\u003csub\u003ei\u003c/sub\u003e is the coordinates of the i-th point, and P\u0026apos; is a point from the already sampled point set. The index corresponding to argmax(D\u003csub\u003ei\u003c/sub\u003e) is chosen as the next sampling point.\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eThis FPS strategy yields a downsampled point cloud along with the corresponding indices of the sampled points. In scenarios involving large datasets or limited computational resources, this sampling method proves instrumental in reducing the data volume while still capturing the essential structure and details of the point cloud. Such efficiency is crucial in point cloud processing for fabric analysis. The downsampled point clouds effectively preserve the structural integrity and detailed textural information of the fabric surfaces, essential for accurately characterizing wrinkle patterns, as illustrated in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e\n \u003ch2\u003e2.3.3. Point cloud normalization\u003c/h2\u003e\n \u003cp\u003eBefore processing point cloud data, certain preprocessing steps are necessary, among which normalization is a key component. Normalization ensures that data of different scales and units can be correctly interpreted and compared.\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eFirstly, the point cloud data are normalized by calculating the mean and standard deviation using the following formulas, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{p}_{i}\\)\u003c/span\u003e\u003c/span\u003e represents each point and N is the total number of points:\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003ctable id=\"Tabc\" style=\"width: 360px;\" border=\"1\"\u003e\n \u003cthead\u003e\n \u003ctr style=\"height: 38.0531px;\"\u003e\n \u003cth style=\"width: 325px; height: 38.0531px;\" align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:mean=\\frac{\\sum\\:{p}_{i}}{N}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/th\u003e\n \u003cth style=\"width: 15.713px; height: 38.0531px;\" colspan=\"2\" align=\"left\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/th\u003e\n \u003cth style=\"width: 10px; height: 38.0531px;\" colspan=\"1\" align=\"left\"\u003e\u0026nbsp;(2)\u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr style=\"height: 48px;\"\u003e\n \u003ctd style=\"width: 340.713px; height: 48px;\" colspan=\"3\" align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:std=\\sqrt{\\frac{\\sum\\:({p}_{i}-mean{)}^{2}}{N}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px; height: 48px;\" align=\"left\"\u003e\n \u003cp\u003e(3)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eThen, the point cloud data are processed using zero-mean normalization (also known as standardization) to ensure the normalized data has a mean of 0 and a standard deviation of 1. This standardization is crucial as it facilitates a uniform platform for analyzing intricate wrinkle characteristics by mitigating scale disparities across different scans. The formula is as follows:\u003c/p\u003e\n \u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\n \u003cdiv id=\"FileID_Equ2\" class=\"mathdisplay\"\u003e$$\\:points\\_ormalized=\\frac{points-mean}{std}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eNext, the normals of the point cloud are processed. The maximum Euclidean norm of all the normals is calculated using the following formula, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{p}_{i}\\)\u003c/span\u003e\u003c/span\u003e is each normal vector:\u003c/p\u003e\n \u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\n \u003cdiv id=\"FileID_Equ3\" class=\"mathdisplay\"\u003e$$\\:max\\_norm=max\\left(\\sqrt{\\sum\\:{p}_{i}^{2}}\\right)$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eTo achieve consistent analysis quality across all normals, they are normalized using this maximum norm by dividing each normal vector by the maximum norm:\u003c/p\u003e\n \u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\n \u003cdiv id=\"FileID_Equ4\" class=\"mathdisplay\"\u003e$$\\:normals\\_normalized=\\frac{normals}{max\\_norm}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eFinally, the normalized point cloud data and normals are recombined to form a complete point cloud dataset. The normalization process ensures that the scale of the point cloud data and normal values is consistent, which is crucial for subsequent analysis and processing aimed at wrinkle characterization. As illustrated in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e, blue indicates points closest to the sample, and yellow marks those farthest from the reference plane of the sample. Due to changes in the reference plane caused by downsampling and normalization\u0026mdash;particularly when downsampling reduces data to a tenth of its original scale\u0026mdash;the appearance of the point cloud might alter when viewed from a fixed angle. Nonetheless, it is evident that the essential characteristics of the fabric wrinkles are preserved, allowing for accurate analysis of wrinkle patterns and textures after these processing steps.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e\n \u003ch2\u003e2.3.4. Data Augmentation\u003c/h2\u003e\n \u003cp\u003eDue to the limited scale of the original dataset, deep learning models may struggle to learn the underlying patterns and complexities of fabric wrinkles, thus affecting their ability to generalize these features[\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e]. Small-scale datasets can lead to overfitting and poor performance on unseen fabric data[\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e], while imbalanced datasets may further impair the models\u0026apos; ability to recognize certain types of wrinkles due to insufficient data [\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e]. To address these issues, four methods were employed to enhance the dataset\u0026apos;s robustness in representing diverse wrinkle characteristics.\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eA) Rotation\u003c/em\u003e:\u003c/p\u003e\n \u003cp\u003eRotation is a crucial operation in point cloud data processing. It involves converting the rotation angle from degrees to radians and performing a rotation using a 3x3 rotation matrix. In this study, 8 rotations were performed on the fabric point cloud using a 45-degree rotation angle to ensure comprehensive analysis from multiple perspectives, which is vital for understanding the varied wrinkle patterns on the fabric surface.\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eB) Translation\u003c/em\u003e\u003c/p\u003e\n \u003cp\u003eTranslation is a geometric transformation that allows for the movement of the entire point cloud along specified directions without changing their shape or orientation. This operation is important for aligning and comparing different sections of fabric when analyzing wrinkle characteristics.\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eC) Noise\u003c/em\u003e\u003c/p\u003e\n \u003cp\u003eAdding random noise to the data helps assess the robustness of the point cloud processing approach specifically for fabric wrinkle analysis. Gaussian noise, with the same shape as the point cloud, is generated and added to the original points. The surface normals are unaffected by the noise to maintain the original surface orientation. In this study, five Gaussian noises with random frequencies were generated and two of them were randomly applied to the point cloud for each iteration, as shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e (d). This operation is crucial for ensuring the fabric\u0026apos;s wrinkle characteristics can be reliably analyzed under various noise conditions.\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eD) Scaling\u003c/em\u003e\u003c/p\u003e\n \u003cp\u003eScaling allows for changing the size of objects or adjusting the distance between objects and observers, which is essential for analyzing how these variations affect the visibility and characteristics of fabric wrinkles. Each point in the point cloud is multiplied by a scaling factor, while the surface normals remain unchanged to maintain accurate orientation of the fabric surfaces. In this study, five scaling factors were selected: 1.2, 1.1, 1.0, 0.9, and 0.8, and two of them were randomly applied to the point cloud for each iteration, as demonstrated in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e (e). This operation is crucial for examining the effects of scale on fabric wrinkles, ensuring that the analysis captures the true nature of wrinkle patterns under different conditions.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n \u003ch2\u003e2.4. Evaluation models\u003c/h2\u003e\n \u003cp\u003eEvaluation neural networks for processing fabric wrinkle point cloud data possess several key characteristics:\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eA) Handling unordered data\u003c/em\u003e:\u003c/p\u003e\n \u003cp\u003ePoint cloud data is unordered, allowing for arbitrary point arrangements without affecting the overall shape information. Neural networks handle this through pooling operations using symmetric functions such as maximum or average value[\u003cspan class=\"CitationRef\"\u003e28\u003c/span\u003e]\u0026ndash;[\u003cspan class=\"CitationRef\"\u003e30\u003c/span\u003e].\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eB) Invariance to rotation and scaling\u003c/em\u003e:\u003c/p\u003e\n \u003cp\u003ePoint cloud data can vary in rotation and scaling. Neural networks achieve invariance to transformations by incorporating alignment or normalization steps within the network [\u003cspan class=\"CitationRef\"\u003e31\u003c/span\u003e]\u0026ndash;[\u003cspan class=\"CitationRef\"\u003e33\u003c/span\u003e].\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eC) Integration of Local and Global Features\u003c/em\u003e:\u003c/p\u003e\n \u003cp\u003ePoint cloud data contains both local and global shape information, as well as multi-scale feature learning [\u003cspan class=\"CitationRef\"\u003e34\u003c/span\u003e].\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eD) Efficient Spatial Search and Computation\u003c/em\u003e:\u003c/p\u003e\n \u003cp\u003ePoint cloud data consists of a large number of points, requiring efficient spatial search and computation. This is achieved through optimization algorithms and data structures such as K-D trees and octrees[\u003cspan class=\"CitationRef\"\u003e35\u003c/span\u003e]\u0026ndash;[\u003cspan class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e\n \u003cp\u003eNeural networks for point cloud data processing excel in various applications, including 3D object recognition, semantic segmentation, and point cloud generation.\u003c/p\u003e\n \u003cp\u003eThe PointNet series of networks handle point cloud data directly by utilizing symmetric functions for global feature learning and employing feature transformation networks for alignment and normalization. This approach ensures invariance to point cloud arrangement, rotation, and translation, while also adapting well to variations in point cloud size and density. In this study, the PointNet series of networks is used to evaluate the surface smoothness of textile materials after preprocessing the point cloud data.\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e\n \u003ch2\u003e2.4.1. PointNet architecture\u003c/h2\u003e\n \u003cp\u003ePointNet[\u003cspan class=\"CitationRef\"\u003e39\u003c/span\u003e] is a neural network designed to directly process point cloud data. Its primary characteristic lies in its ability to extract useful features directly from raw point cloud data, thereby eliminating the need for complex data preprocessing or transformation. The network structure of PointNet primarily consists of two parts: the Input Transform and the Evaluation Network, as shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e9\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eA) Input Transform\u003c/em\u003e:\u003c/p\u003e\n \u003cp\u003eThis is a miniaturized version of PointNet, with the objective of learning an alignment matrix to align the input point cloud to a normalized coordinate space. This process ensures the network\u0026apos;s invariance to the rotation and translation of the input point cloud.\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eB) Evaluation Network\u003c/em\u003e:\u003c/p\u003e\n \u003cp\u003eThe Evaluation Network forms the main body of PointNet. It initially extracts the features of each point through a series of Multi-Layer Perceptrons (MLP), then obtains global features through a max pooling layer. The max pooling layer ensures the network\u0026apos;s invariance to the order of the input points. Finally, the global features are processed through another series of MLPs to obtain the final evaluation result.\u003c/p\u003e\n \u003cp\u003eDuring the implementation process, PointNet incorporates a T-Net at each stage of feature extraction. This enhances the network\u0026apos;s invariance to the rotation and scaling of the input point cloud.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e\n \u003ch2\u003e2.4.2. PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;architecture\u003c/h2\u003e\n \u003cp\u003ePointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;is a network that builds upon the foundation of PointNet to address some of the challenges encountered when dealing with complex structures and large-scale point cloud data, as shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e10\u003c/span\u003e. The main improvements in PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;include the introduction of hierarchical feature learning and local structure modeling of point clouds.\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eA) Hierarchical Point Set Feature Learning\u003c/em\u003e:\u003c/p\u003e\n \u003cp\u003eInstead of directly performing feature learning on the entire point cloud, PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;adopts a different approach. It first divides the point cloud into multiple smaller regions and applies PointNet for feature learning on each of these regions individually. This is achieved by sampling and aggregating the neighborhood of each point. The hierarchical feature learning method enables PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;to better handle large-scale and complex point cloud data.\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eB) Evaluation Network\u003c/em\u003e:\u003c/p\u003e\n \u003cp\u003eWhen computing local features, PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;considers not only the characteristics of each point itself but also the characteristics of its neighboring points. The local features obtained from these regions are then processed using Multi-Layer Perceptrons (MLPs) to derive global features. This approach captures the local structural information of the point cloud, thereby improving the performance of the model.\u003c/p\u003e\n \u003cp\u003eCompared to PointNet, PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;exhibits superior performance in handling complex structures and large-scale point cloud data. However, it also comes with higher computational complexity and storage requirements.\u003c/p\u003e\n \u003cp\u003eIn general, PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;typically performs a single random farthest point sampling and sets a grouping radius to extract PointNet features from the grouped point clouds. However, the Multi-Scale Grouping (MSG) approach utilizes a different strategy, as shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e11\u003c/span\u003e. It also conducts a random farthest point sampling, but with multiple grouping radii specified for each group. Subsequently, PointNet features are extracted individually for each group. Finally, the point cloud features extracted at different grouping radii are concatenated to obtain the final features output by the Set Abstraction (SA) module.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"3. Results and discussion","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1. Experimental environment\u003c/h2\u003e\n \u003cp\u003eThe configuration for the network training and testing environment is as follows: Processor: 12th Gen Intel(R) Core(TM) i7\u0026ndash;12700K; Memory: 32 GB; Operating system: Windows 11 64-bit; GPU: NVIDIA GeForce RTX 3080; Programming platform: Python 3.9; Deep learning framework: PyTorch 1.6.0.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2. Dataset Prepared\u003c/h2\u003e\n \u003cp\u003eIn the experiment, the point cloud dataset for each type of fabric surface wrinkle is divided into a training set and a test set at a 4:1 ratio, as shown in Table\u0026nbsp;3. A total of 44 original point clouds of 12 types of fabric surface wrinkle were collected, as shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e12\u003c/span\u003e. Each original point cloud is augmented to generate 40 augmented point clouds, resulting in a total of 1760 samples, as shown in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e13\u003c/span\u003e. From each group of 40 augmented point clouds, 32 point clouds are randomly selected, resulting in a total of 1408 point clouds used as training samples. The remaining 8 augmented point clouds from each original point cloud are used as test samples, resulting in a total of 352 test samples.\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eTable.3 Fabric Surface Wrinkle point cloud dataset\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tabd\" border=\"1\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eType\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNumber of training\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNumber of testing\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWR-1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e128\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWR-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e256\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e64\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWR-3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e320\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e80\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWR-4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e448\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e112\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWR-5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e256\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e64\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\n \u003ch2\u003e3.3. Evaluation results and analysis\u003c/h2\u003e\n \u003cdiv id=\"Sec17\" class=\"Section3\"\u003e\n \u003ch2\u003e3.3.1. PointNet\u003c/h2\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e14\u003c/span\u003e illustrates the training progress of the PointNet transfer learning model with different configurations: (a) PointNet (without normal) and (b) PointNet. From the experimental results in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e14\u003c/span\u003ea, it is observed that for PointNet (without normal), the accuracy measures start to ascend notably after 110 epochs. However, even reaching up to 200 epochs, the model does not achieve a stable or high accuracy level, failing to surpass 80%. This highlights the initial slow and unstable learning process when normal vectors are not included, reflecting PointNet\u0026apos;s challenge in capturing the nuanced features of fabric wrinkles without sufficient directional information provided by normals.\u003c/p\u003e\n \u003cp\u003eIn contrast, Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e14\u003c/span\u003eb depicts the training progress for the standard PointNet model. Here, the model demonstrates a more robust performance. While the early epochs exhibit some variability, the accuracy quickly approaches optimal levels, demonstrating the effectiveness of including normal vectors in capturing fabric wrinkles more accurately and efficiently. The graph shows a more rapid and stable convergence compared to the variant without normal vectors, indicating that the normal vector inclusion significantly aids in the model\u0026apos;s ability to discern detailed surface features and improves overall learning effectiveness.\u003c/p\u003e\n \u003cp\u003eThese observations from Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e14\u003c/span\u003e are supported by the accuracy data in Table\u0026nbsp;4 and Table\u0026nbsp;5, where PointNet with normal vectors consistently shows superior performance in training and evaluation tasks compared to its counterpart without normal vectors.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003c/div\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec18\" class=\"Section3\"\u003e\n \u003ch2\u003e3.3.2. PointNet++\u003c/h2\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e15\u003c/span\u003e showcases the training progress of the PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;transfer learning model with comparisons between configurations with and without normal vectors. In Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e15\u003c/span\u003ea, when normal vectors are included, PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;demonstrates a rapid learning curve, achieving over 95% training accuracy within just 50 epochs. This is a significant improvement over the standard PointNet model, which reaches similar levels of training accuracy only after approximately 100 epochs as depicted in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e14\u003c/span\u003eb. The enhanced training efficiency of PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;is attributed to its ability to more effectively identify fabric wrinkle areas by incorporating additional components focusing on local features. However, with the inclusion of normal vectors, PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;may experience fluctuations in accuracy during the training process, particularly in test accuracy and class accuracy, which oscillate between 80%-95% as iterations proceed. These fluctuations are likely due to the model\u0026apos;s heightened sensitivity to the subtle differences between wrinkled and non-wrinkled areas in the presence of normal vectors. The data shows that while PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;can quickly learn and distinguish different wrinkle grades, the finer details captured by normal vectors introduce variability during the early training phases.\u003c/p\u003e\n \u003cp\u003eIn contrast, Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e15\u003c/span\u003eb illustrates the training progress of PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;without normal vectors. Notably, even while retaining a higher level of training efficiency, the model demonstrates reduced fluctuations in test accuracy and class accuracy, with these metrics becoming more stable after about 125 epochs. The removal of normal vectors seems to mitigate the model\u0026apos;s sensitivity to minute surface variations, leading to a smoother and more predictable learning process while still maintaining a high level of accuracy.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec19\" class=\"Section3\"\u003e\n \u003ch2\u003e3.3.3. PointNet++_MSG\u003c/h2\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e16\u003c/span\u003e depicts the training progress of the PointNet++_MSG transfer learning model with and without normal vectors, demonstrating its sensitivity to local features and subsequent impact on training stability and accuracy. Figure\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e16\u003c/span\u003ea, which illustrates the PointNet++_MSG model with normal vectors, shows more pronounced fluctuations in accuracy, oscillating between 50%-90% within 200 epochs, indicating a less stable training process compared to the other models. These fluctuations are attributed to the model\u0026apos;s high sensitivity to local information provided by normal vectors, leading to significant changes in accuracy with each iteration, especially when the dataset is complex or contains detailed wrinkle features.\u003c/p\u003e\n \u003cp\u003eIn contrast, Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e16\u003c/span\u003eb shows the training progress of the PointNet++_MSG model after the removal of normal vectors. Here, the model demonstrates delayed training accuracy, reaching around 92% by the 25th epoch, compared to the 95% achieved by PointNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;at the same epoch count. However, after 50 epochs, PointNet++_MSG (without normal) exhibits a trend towards stability in both test accuracy and class accuracy, achieving a more stable and reliable performance sooner than PointNet++, which only begins to stabilize after about 125 epochs. This indicates that while the initial learning may be slightly slower without normal vectors, the overall training process becomes more stable and efficient, reducing the wide accuracy fluctuations seen with normal vectors.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eTable.4 Training and Testing Accuracy of PointNet Series Networks\u003c/strong\u003e\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003ctable id=\"Tabh\" border=\"1\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eItem\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTrain Accuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTest Accuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eBest Accuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9688\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet(without normal)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.5355\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.6818\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7121\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet++\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9950\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7045\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.6750\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet++(without normal)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9950\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet++_MSG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9950\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9290\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9245\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet++_MSG(without normal)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9979\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003e\u003cstrong\u003eTable.5 Evaluation Accuracy of PointNet Series Networks\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tabi\" border=\"1\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eItem\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eEvaluate Accuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eBEA (Best Evaluate Accuracy)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet(without normal)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.6989\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7339\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet++\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9773\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9816\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet++_MSG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9943\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9938\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eIn order to confirm the validity of the methodology proposed in this paper, a comparative analysis was carried out with the methodology used in previous studies, which is shown in Table\u0026nbsp;6. The results show that, in general, the methodology used in this study is superior in terms of accuracy and reliability.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eTable.6 Comparison of evaluation accuracy of different methods\u003c/strong\u003e\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n \u003ctable id=\"Tabj\" border=\"1\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTrain Accuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTest Accuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eBest Accuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eEvaluate Accuracy\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eBEA\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSVM[\u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.3984\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.4062\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.4062\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.4062\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.4062\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOptimized SVM[\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7195\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7131\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7195\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7131\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7131\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eKNN[\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9247\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.8920\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9247\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.8920\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.8920\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet(without normal)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.5355\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.6818\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7121\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.6989\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7339\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet++\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9950\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.7045\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.6750\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9773\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9816\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePointNet++_MSG\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9950\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9290\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e0.9245\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9943\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.9938\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"4. Conclusions","content":"\u003cp\u003eUtilizing advanced 3D point cloud data and machine learning algorithms, this research achieves precise evaluation of fabric surface wrinkles. A robust dataset was assembled from 44 three-dimensional point cloud samples spanning 12 fabric types and 5 AATCC-WR grades with the aid of the EinScan-SP scanner, enhanced through techniques such as rotation, translation, noise addition, and scaling.\u003c/p\u003e \u003cp\u003eEvaluations employing PointNet, PointNet++, and PointNet++_MSG architectures demonstrate that these methods, especially when augmented by normal vectors, significantly outperform traditional wrinkle assessment techniques. The use of optimized multi-scale grouping in PointNet++_MSG effectively minimizes interference from point normals, leading to accurate assessments.\u003c/p\u003e \u003cp\u003eThese advancements underscore the potential of machine learning in transforming textile quality control by setting new standards for precision and reliability in fabric assessment. The non-contact nature and reduced sensitivity to fabric variations further enhance its applicability across the textile industry, marking a significant step forward in integrating computational technology into traditional practices.\u003c/p\u003e"},{"header":"Declarations","content":" \u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eYuan completed the primary research and drafted the main content. Xin formulated the research direction. Newton reviewed the manuscript for English grammar. Yuan and Shen created the dataset. Zhang refined the research content. All authors reviewed the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThis project was funded by the National Natural Science Foundation of China (Grant No. 11702169).\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data supporting the findings of this study are available upon reasonable request from the authors. Code and dataset links in this article: https://github.com/YuanZhijie/FabricPointNet.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eM. S. Brown, B. Ashley, and A. Koh, \u0026ldquo;Wearable technology for chronic wound monitoring: current dressings, advancements, and future prospects,\u0026rdquo; Frontiers in bioengineering and biotechnology, vol. 6, p. 47, 2018, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fbioe.2018.00047\u003c/span\u003e\u003cspan address=\"10.3389/fbioe.2018.00047\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA. Gurarda, \u0026ldquo;Seam performance of garments,\u0026rdquo; in Textile Manufacturing Processes, IntechOpen, 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eN. Shen, A. Samanta, H. Ding, and W. W. Cai, \u0026ldquo;Simulating microstructure evolution of battery tabs during ultrasonic welding,\u0026rdquo; Journal of Manufacturing Processes, vol. 23, pp. 306\u0026ndash;314, 2016.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eR. Narain, A. Samii, and J. F. O\u0026rsquo;brien, \u0026ldquo;Adaptive anisotropic remeshing for cloth simulation,\u0026rdquo; ACM transactions on graphics (TOG), vol. 31, no. 6, pp. 1\u0026ndash;10, 2012.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eX. Chen, M. Yu, and Y. Song, \u0026ldquo;Optimized Seam-Driven Image Stitching Method Based on Scene Depth Information,\u0026rdquo; Electronics, vol. 11, no. 12, p. 1876, 2022, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/electronics11121876\u003c/span\u003e\u003cspan address=\"10.3390/electronics11121876\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM. Shiloh, \u0026ldquo;The evaluation of seam-puckering,\u0026rdquo; 1971, doi: 10.1080/00405007108630042.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. Galuszynski, \u0026ldquo;Effect of fabric structure on fabric resistance to needle piercing,\u0026rdquo; 1986, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/004051758605600511\u003c/span\u003e\u003cspan address=\"10.1177/004051758605600511\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eF. Sun, M. Guo, X. Hu, L. Wang, and W. Gao, \u0026ldquo;Analysis of curve parameters to characterize multidirectional fabric wrinkling by a double extraction method,\u0026rdquo; Textile Research Journal, vol. 89, no. 15, pp. 2973\u0026ndash;2982, Aug. 2019, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/0040517518805372\u003c/span\u003e\u003cspan address=\"10.1177/0040517518805372\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eY. Lu, X. Hu, F. Sun, F. Peng, and W. Gao, \u0026ldquo;Determination of optimal system parameters to characterize the wrinkle recovery of fabrics by an integrated shape retention evaluation system,\u0026rdquo; Textile Research Journal, vol. 90, no. 1, pp. 91\u0026ndash;100, Jan. 2020, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/0040517519858770\u003c/span\u003e\u003cspan address=\"10.1177/0040517519858770\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZ. Zhou, R. Zhang, J. Zhang, Y. Wang, Z. Zhu, and C. Liu, \u0026ldquo;Fabric wrinkle level classification via online sequential extreme learning machine based on improved sine cosine algorithm,\u0026rdquo; Textile Research Journal, vol. 90, no. 17\u0026ndash;18, pp. 2007\u0026ndash;2021, Sep. 2020, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/0040517520908072\u003c/span\u003e\u003cspan address=\"10.1177/0040517520908072\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZ. Zhou, Z. Ma, Y. Wang, and Z. Zhu, \u0026ldquo;Fabric wrinkle rating model based on ResNet18 and optimized random vector functional-link network,\u0026rdquo; Textile Research Journal, vol. 93, no. 1\u0026ndash;2, pp. 172\u0026ndash;193, Jan. 2023, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/00405175221117614\u003c/span\u003e\u003cspan address=\"10.1177/00405175221117614\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eR. Furferi, L. Governi, and Y. Volpe, \u0026ldquo;Machine Vision-Based Pilling Assessment: A Review,\u0026rdquo; Journal of Engineered Fibers and Fabrics, vol. 10, no. 3, p. 155892501501000, Sep. 2015, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/155892501501000320\u003c/span\u003e\u003cspan address=\"10.1177/155892501501000320\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eP. Yildirim, D. Birant, and T. Alpyildiz, \u0026ldquo;Data mining and machine learning in textile industry,\u0026rdquo; WIREs Data Min \u0026amp; Knowl, vol. 8, no. 1, p. e1228, Jan. 2018, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/widm.1228\u003c/span\u003e\u003cspan address=\"10.1002/widm.1228\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA. M. Bertoldi and D. L. Munden, \u0026ldquo;The effects of sewing variables on fabric pucker,\u0026rdquo; Cloth. Res. J, vol. 2, no. 2, p. 81, 1974.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eL. L. Chen, \u0026ldquo;Objective evaluation method for sewing flatness of worsted wool fabrics,\u0026rdquo; J Text Res, vol. 3, pp. 120\u0026ndash;125, 2018.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM. Mohri, S. H. Ravandi, and M. Youssefi, \u0026ldquo;Objective evaluation of wrinkled fabric using radon transform,\u0026rdquo; Journal of the Textile Institute, vol. 96, no. 6, pp. 365\u0026ndash;370, 2005, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1533/joti.2004.0066\u003c/span\u003e\u003cspan address=\"10.1533/joti.2004.0066\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eX. B. Yang and X. B. Huang, \u0026ldquo;Evaluating fabric wrinkle degree with a photometric stereo method,\u0026rdquo; Textile Research Journal, vol. 73, no. 5, pp. 451\u0026ndash;454, 2003, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/004051750307300513\u003c/span\u003e\u003cspan address=\"10.1177/004051750307300513\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJingjing Sun, Ming Yao, Bugao Xu, and P. Bel, \u0026ldquo;Fabric wrinkle characterization and classification using modified wavelet coefficients and support-vector-machine classifiers,\u0026rdquo; Textile Research Journal, vol. 81, no. 9, pp. 902\u0026ndash;913, Jun. 2011, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/0040517510391702\u003c/span\u003e\u003cspan address=\"10.1177/0040517510391702\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eN. Deng, Y. Wang, B. Xin, and W. Wang, \u0026ldquo;A novel objective wrinkle evaluation method for printed fabrics based on multi-view stereo algorithm,\u0026rdquo; The Journal of The Textile Institute, vol. 113, no. 3, pp. 367\u0026ndash;377, Mar. 2022, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/00405000.2021.1882071\u003c/span\u003e\u003cspan address=\"10.1080/00405000.2021.1882071\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eY. Wang, N. Deng, and B. Xin, \u0026ldquo;Investigation of 3D surface profile reconstruction technology for automatic evaluation of fabric smoothness appearance,\u0026rdquo; Measurement, vol. 166, p. 108264, Dec. 2020, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.measurement.2020.108264\u003c/span\u003e\u003cspan address=\"10.1016/j.measurement.2020.108264\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eB. Yang, N. Haala, and Z. Dong, \u0026ldquo;Progress and perspectives of point cloud intelligence,\u0026rdquo; Geo-spatial Information Science, pp. 1\u0026ndash;17, 2023, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/10095020.2023.2175478\u003c/span\u003e\u003cspan address=\"10.1080/10095020.2023.2175478\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eB. Shen, L. Li, X. Hu, S. Guo, J. Huang, and Z. Liang, \u0026ldquo;Point cloud upsampling generative adversarial network based on residual multi-scale off-set attention,\u0026rdquo; Virtual Reality \u0026amp; Intelligent Hardware, vol. 5, no. 1, pp. 81\u0026ndash;91, 2023, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.vrih.2022.08.016\u003c/span\u003e\u003cspan address=\"10.1016/j.vrih.2022.08.016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZ. Zhou, Y. Hu, Z. Zhu, and Y. Wang, \u0026ldquo;Fabric Wrinkle Objective Evaluation Model with Random Vector Function Link Based on Optimized Artificial Hummingbird Algorithm,\u0026rdquo; J. Nat. Fibers, vol. 20, no. 1, p. 2163026, Dec. 2023, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/15440478.2022.2163026\u003c/span\u003e\u003cspan address=\"10.1080/15440478.2022.2163026\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eY. Liu, X. Sun, W. Shao, and Y. Yuan, \u0026ldquo;S2ANet: Combining local spectral and spatial point grouping for point cloud processing,\u0026rdquo; Virtual Reality \u0026amp; Intelligent Hardware, vol. 6, no. 4, pp. 267\u0026ndash;279, 2024, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.vrih.2023.06.005\u003c/span\u003e\u003cspan address=\"10.1016/j.vrih.2023.06.005\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM. Oleynik, A. Kugic, Z. Kas\u0026aacute;č, and M. Kreuzthaler, \u0026ldquo;Evaluating shallow and deep learning strategies for the 2018 n2c2 shared task on clinical text classification,\u0026rdquo; Journal of the American Medical Informatics Association, vol. 26, no. 11, pp. 1247\u0026ndash;1254, 2019, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/jamia/ocz149\u003c/span\u003e\u003cspan address=\"10.1093/jamia/ocz149\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. Harrer, P. Shah, B. Antony, and J. Hu, \u0026ldquo;Artificial intelligence for clinical trial design,\u0026rdquo; Trends in pharmacological sciences, vol. 40, no. 8, pp. 577\u0026ndash;591, 2019, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.tips.2019.05.005\u003c/span\u003e\u003cspan address=\"10.1016/j.tips.2019.05.005\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eX. Zhang, J. Chen, Y. Zhou, L. Han, and J. Lin, \u0026ldquo;A multiple-layer representation learning model for network-based attack detection,\u0026rdquo; IEEE Access, vol. 7, pp. 91992\u0026ndash;92008, 2019, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ACCESS.2019.2927465\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2019.2927465\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eW. Zhang, S. Su, B. Wang, Q. Hong, and L. Sun, \u0026ldquo;Local k-NNs pattern in Omni-Direction graph convolution neural network for 3D point clouds,\u0026rdquo; Neurocomputing, vol. 413, pp. 487\u0026ndash;498, 2020, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neucom.2020.06.095\u003c/span\u003e\u003cspan address=\"10.1016/j.neucom.2020.06.095\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM. D. Phung, C. H. Quach, D. T. Chu, N. Q. Nguyen, T. H. Dinh, and Q. P. Ha, \u0026ldquo;Automatic interpretation of unordered point cloud data for uav navigation in construction,\u0026rdquo; in 2016 14th International Conference on Control, Automation, Robotics and Vision (ICARCV), IEEE, 2016, pp. 1\u0026ndash;6. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ICARCV.2016.7838683\u003c/span\u003e\u003cspan address=\"10.1109/ICARCV.2016.7838683\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eW. Liu, J. Sun, W. Li, T. Hu, and P. Wang, \u0026ldquo;Deep learning on point clouds and its application: A survey,\u0026rdquo; Sensors, vol. 19, no. 19, p. 4188, 2019, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/s19194188\u003c/span\u003e\u003cspan address=\"10.3390/s19194188\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eK. Al-Manasir and C. S. Fraser, \u0026ldquo;Registration of terrestrial laser scanner data using imagery,\u0026rdquo; The Photogrammetric Record, vol. 21, no. 115, pp. 255\u0026ndash;268, 2006, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/j.1477-9730.2006.00379.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1477-9730.2006.00379.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA. Zaganidis, L. Sun, T. Duckett, and G. Cielniak, \u0026ldquo;Integrating deep semantic segmentation into 3-d point cloud registration,\u0026rdquo; IEEE Robotics and Automation Letters, vol. 3, no. 4, pp. 2942\u0026ndash;2949, 2018, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/LRA.2018.2848308\u003c/span\u003e\u003cspan address=\"10.1109/LRA.2018.2848308\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. A. Eslami et al., \u0026ldquo;Neural scene representation and rendering,\u0026rdquo; Science, vol. 360, no. 6394, pp. 1204\u0026ndash;1210, 2018, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1126/science.aar6170\u003c/span\u003e\u003cspan address=\"10.1126/science.aar6170\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eY. Wang, Y. Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon, \u0026ldquo;Dynamic graph cnn for learning on point clouds,\u0026rdquo; Acm Transactions On Graphics (tog), vol. 38, no. 5, pp. 1\u0026ndash;12, 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eF. Wang and Z. Zhao, \u0026ldquo;A survey of iterative closest point algorithm,\u0026rdquo; in 2017 Chinese Automation Congress (CAC), IEEE, 2017, pp. 4395\u0026ndash;4399. doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/CAC.2017.8243553\u003c/span\u003e\u003cspan address=\"10.1109/CAC.2017.8243553\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZ. Chen et al., \u0026ldquo;An artificial bee bare-bone hunger games search for global optimization and high-dimensional feature selection,\u0026rdquo; Iscience, vol. 26, no. 5, 2023, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.isci.2023.106679\u003c/span\u003e\u003cspan address=\"10.1016/j.isci.2023.106679\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eX. Shi, T. Liu, and X. Han, \u0026ldquo;Improved Iterative Closest Point (ICP) 3D point cloud registration algorithm based on point cloud filtering and adaptive fireworks for coarse registration,\u0026rdquo; International Journal of Remote Sensing, vol. 41, no. 8, pp. 3197\u0026ndash;3220, 2020, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/01431161.2019.1701211\u003c/span\u003e\u003cspan address=\"10.1080/01431161.2019.1701211\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eC. Wang, M. Ji, J. Wang, W. Wen, T. Li, and Y. Sun, \u0026ldquo;An improved DBSCAN method for LiDAR data segmentation with automatic Eps estimation,\u0026rdquo; Sensors, vol. 19, no. 1, p. 172, 2019, doi: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/s19010172\u003c/span\u003e\u003cspan address=\"10.3390/s19010172\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. Xie, Y. Xu, Z. Zheng, S.-C. Zhu, and Y. N. Wu, \u0026ldquo;Generative pointnet: Deep energy-based learning on unordered point sets for 3d generation, reconstruction and classification,\u0026rdquo; in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 14976\u0026ndash;14985.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"3D Point Cloud Data, Fabric Surface Wrinkle, Quality Control, PointNet + + Network","lastPublishedDoi":"10.21203/rs.3.rs-5220842/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5220842/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study presents a novel, high-precision method for objectively assessing fabric wrinkles utilizing 3D point cloud data and advanced deep learning techniques. Employing the EinScan-SP scanner, we generated a comprehensive dataset comprising 44 samples from 12 distinct fabric types. The intricate surface topologies of these fabrics were analyzed using PointNet, PointNet++, and PointNet++_MSG architectures, demonstrating their sensitivity to subtle wrinkle details. Our non-contact, automated approach significantly improves upon traditional wrinkle assessment techniques, offering an accurate and reliable means of quality control in the textile industry. The methodology's effectiveness was validated through comparative analysis, showcasing its superiority in terms of accuracy and repeatability. This study establishes a new benchmark for the precise evaluation of fabric surface characteristics, facilitating advancements in textile quality standards. The code and datasets are publicly available at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/YuanZhijie/FabricPointNet\u003c/span\u003e\u003cspan address=\"https://github.com/YuanZhijie/FabricPointNet\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e","manuscriptTitle":"Objective and High-Precision Fabric Wrinkle Assessment using 3D Point Cloud Data and Deep Learning Techniques","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-12-23 09:31:31","doi":"10.21203/rs.3.rs-5220842/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"7fa622fa-1ea4-440e-8289-9b6d2f5741c2","owner":[],"postedDate":"December 23rd, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-12-13T14:39:30+00:00","versionOfRecord":[],"versionCreatedAt":"2024-12-23 09:31:31","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5220842","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5220842","identity":"rs-5220842","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}