Semantic Segmentation of Steel Wire Rod Thermal Images for Automated Temperature Measurement

Semantic Segmentation of Steel Wire Rod Thermal Images for Automated Temperature Measurement | 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 Semantic Segmentation of Steel Wire Rod Thermal Images for Automated Temperature Measurement Seok-Kyu Pyo, Dong-Hee Lee, Sung-Jun Hur, Sang-Hyeon Lee, Sung-Jun Lim, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7625115/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 12 Mar, 2026 Read the published version in The International Journal of Advanced Manufacturing Technology → Version 1 posted 5 You are reading this latest preprint version Abstract Hardness is a critical quality attribute of wire rods and is significantly influenced by cooling temperatures. To measure this temperature, infrared line scanners are typically employed to generate thermal images that provide full-length temperature data of wire rods. In these thermal images, some pixels correspond to the rod surface temperature, while others represent the background, which often contains noise. Accurately identifying rod-related pixels is essential, but this remains challenging due to the complex geometry of the wire rods. To address this, we propose a semantic segmentation method using a U-Net architecture to automatically distinguish wire rod pixels from the background. This method improves automation and eliminates the need for manual processing. We validated the approach on wire rods of varying diameters, from heavy-type (> 10 mm) to fine-type (≤ 10 mm). For heavy-type rods, our method achieved an accuracy of 0.9624 and a mean Intersection over Union of 0.8998. For fine-type wire rods, where smaller temperature differences complicate segmentation, we introduced a restoration technique using a sliding window approach, which enhanced segmentation quality. This combined method enables automated full-length temperature measurement across different wire rod types, supporting improved quality control in wire rod manufacturing. wire rod thermal image semantic segmentation convolutional neural 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 1. Introduction A wire rod is a semi-finished steel product that must meet specific requirements regarding its physical properties. Wire rods delivered to customers are further processed into finished products such as wires, springs, bolts, and nuts. Noncompliance with these physical specifications may result in product failure or a reduction in the quality of the finished product. Key physical properties include fatigue strength, tensile strength, yield strength, and hardness (Harste & Wustner, 2004 ). Among these properties, hardness is the most important, particularly for automotive suspension springs. When the hardness does not meet specifications, fatigue strength and ductility typically decrease (Koymatcik et al., 2018 ; Pavlina & Van Tyne, 2008 ; Wei et al., 2014 ). Reduced ductility can lead to damage during spring drawing, while decreased fatigue strength compromises the durability and safety of the suspension system. In extreme cases, product fractures can result in costly rework and reputational damage for wire rod manufacturers. To improve the quality of wire rods, it is necessary to both measure and predict their hardness. Rockwell hardness testing is a conventional measurement method based on specimen sampling (Abraham, 2013 ). However, because wire rods are coiled into long lengths, samples are typically taken only from the ends, limiting the comprehensiveness of the inspection. Furthermore, wire rods are hundreds to thousands of meters long, and hardness can vary significantly depending on the position within the coil (Hwang, 2020 ). Even after inspection, defects may still exist in uninspected sections. This limitation can be mitigated by leveraging historical process data to estimate hardness across the entire wire rod (i.e., full-length hardness). One promising approach for estimating full-length hardness is to identify and model the critical factors that influence it. Among the wire rod production stages, this study focuses on the cooling process, which plays a central role in determining hardness. The final transformation occurs during the cooling phase. Figure 1 illustrates a schematic of the cooling process. In this stage, straight wire rods are cooled to the target temperature using water-cooling boxes and are then coiled into loops via the laying head onto the cooling conveyor. These overlapping loops are cooled with forced air using fans, allowing engineers to achieve the desired microstructure and mechanical properties. The cooling rate is a critical factor affecting hardness (Hwang, 2020 ). Therefore, the cooling rate, along with pre-cooling and post-cooling temperatures, must be measured. Measuring the post-cooling temperature of a wire rod is considerably more challenging than measuring the pre-cooling temperature. In the steel-rolling process, temperature is typically measured using an infrared line scanner (Usamentiaga et al., 2020 ). The thermal image obtained from the line scanner contains pixel-by-pixel temperature values. However, these thermal images do not clearly differentiate between pixels representing the material surface (that is, the wire rod) and those representing the background. Since the background pixels are noisy and can distort temperature readings, identifying the material pixels is essential. The complexity of this task depends on the geometry of the material. Figure 2 illustrates the differences between thermal images of a billet and a wire rod. Billets and steel plates have relatively simple shapes, making it easier to distinguish material regions from the background on a region-by-region basis. Conversely, wire rods processed into coils on a laying head exhibit complex geometries, making it challenging to distinguish material from background. Therefore, engineers often manually collect the temperature values from only a few pixels. This manual approach is time-consuming, yields limited data, and increases the likelihood of human error, leading to higher quality control costs. Semantic segmentation offers a promising solution to this challenge. This technique classifies each pixel in an image based on its associated object or region, allowing for the differentiation of material from background across arbitrarily sized image spaces (Long et al., 2015 ). In this context, identifying wire rod pixels in thermal images can be framed as a binary segmentation problem: classifying pixels as either material or background. A semantic segmentation model trained using a convolutional neural network (CNN) architecture can automatically perform this classification. Semantic segmentation techniques have been applied to general photography and electron microscopy images to distinguish between different materials and microstructures (Che et al., 2024 ). However, to the best of our knowledge, no prior studies have applied segmentation techniques to thermal images of wire rods due to their complex shapes. This study aims to develop a segmentation model capable of distinguishing material pixels from background pixels in thermal images of wire rods. The model development process includes the collection of thermal images, image processing, augmentation, semantic segmentation, model training, and validation. Each collected thermal image was divided into patches to enable effective learning, reflecting the fact that a single wire rod spans a very long length during rolling. Individual filtering criteria were applied to each patch to ensure a clear distinction between material and background. Manual annotation was used to generate the training data. To enhance training performance, the data were augmented by applying rotations and inversions, taking advantage of the overall vertical and horizontal symmetry in each patch. Wire rods can be categorized into two types: heavy and fine, based on their diameters. Identifying material pixels in fine-type rods is more challenging, even visually, due to the reduced temperature contrast between the material and background caused by rapid cooling. To address this issue, we propose a restoration technique based on a sliding window approach, which enhances the segmentation quality for fine wire rods. The remainder of this paper is organized as follows. Section 2 reviews the relevant literature. Section 3 describes the proposed segmentation approach. Section 4 explains the additional procedures implemented to improve segmentation performance for fine wire rods. Section 5 summarizes the proposed method and discusses directions for future research. 2. Literature review This section reviews semantic segmentation methods and their applications in the steel industry and other manufacturing domains. Semantic segmentation is a central research area in computer vision that involves classifying each pixel in an image into a specific category. Unlike image classification, which assigns a label to an entire image, or object detection, which identifies and locates discrete objects within an image, semantic segmentation provides detailed, pixel-level classification. A foundational model in this field is the Fully Convolutional Network (FCN), which laid the groundwork for subsequent models such as U-Net and DeepLabV3+ (Yuan et al., 2021 ). U-Net is the most cited model in semantic segmentation, with over 30,000 citations of its original paper (Asthana & Byeon, 2024 ). The name "U-Net" derives from its symmetric encoder-decoder architecture, which resembles the letter “U.” The original U-Net model was designed for biomedical image segmentation, specifically for identifying cells (Ronneberger et al., 2015 ). In the steel industry, semantic segmentation has been primarily utilized for classifying steel scrap, analyzing steel microstructure, and detecting defects during heat treatment. For example, Daigo et al. ( 2023 ) applied semantic segmentation to images of steel scrap to classify different types without direct thickness or diameter measurements. Since different scrap types exhibit varying scales and textures, FCNs often perform poorly due to their limited resolution in deeper layers, which is significantly lower than their resolution in earlier layers. This limitation makes it difficult to detect both fine and coarse features. To address this issue, they proposed a Pyramid Scene Parsing Network (PSPNet), which integrates multi-scale feature representations through a pyramid pooling module. Laub et al. ( 2022 ),Thomas et al. ( 2020 ), andXie et al. ( 2023 ) applied semantic segmentation to classify steel microstructures. BothLaub et al. ( 2022 ) andThomas et al. ( 2020 ) employed U-Net models and performed data augmentation techniques such as rotation and inversion.Laub et al. ( 2022 ) further selected DenseNet as the encoder backbone through empirical evaluation.Xie et al. ( 2023 ) proposed an enhanced Atrous Spatial Pyramid Pooling-FCN to identify multiphase microstructures in steel automatically, and conducted comparative experiments across several networks, including FCN, DeepLab v3+, U-Net, Enet, and PSPNet. Morales-Cervantes et al. ( 2023 ) applied semantic segmentation using CNNs to thermograms for detecting oxidation and decarburization defects during steel heat treatment. Infrared images were obtained during the heating and cooling of a cylindrical steel sample measuring 26 mm in length. The researchers manually labeled each pixel of the images as either background or sample and trained a SegNet model for segmentation. This model achieved a mean accuracy of 92.35% and a mean Intersection over Union (mIoU) of 85.77%. This study shares similarities with our work in that both studies segment thermographic images of steel to distinguish background from material regions. However, our approach differs in several key aspects: we address wire rods that are large, thin, and geometrically complex, and our pipeline includes additional tasks such as image patch segmentation, color filtering, and a restoration model to enhance segmentation quality. Semantic segmentation of thermal images has also been explored in other manufacturing sectors. Xu et al. ( 2018 ) applied DeepLab to segment infrared thermographic images of aluminum electrolysis cells to distinguish between the electrolyte and surface floaters—an important step for accurate temperature measurement. The irregular shapes and inconsistent distribution of the electrolyte posed significant segmentation challenges. Similarly, Lema et al. ( 2023 ), Pedrayes et al. ( 2022 ) applied semantic segmentation to thermography as a non-destructive technique for detecting subsurface defects in carbon fibers. They employed post-processing methods, such as principal component tomography (PCT), in conjunction with U-Net and DeepLabV3 + models to enhance segmentation performance. Overall, our review indicates that semantic segmentation can be successfully applied to thermal images using various deep learning architectures such as FCN, U-Net, and DeepLabV3+. Various preprocessing strategies, including patch-based segmentation and filtering, as well as postprocessing techniques like PCT, have been used to improve model performance. However, no prior studies have applied segmentation techniques to thermal images of wire rods—a domain characterized by complex geometries, irregular placements, and temperature distributions that critically impact product quality. Therefore, applying semantic segmentation to this context is both novel and appropriate. Additionally, existing research has not adequately addressed scenarios in which the temperature contrast between the wire rod and background is minimal, such as in fine-type wire rods. This low contrast presents challenges for conventional preprocessing methods. To resolve this issue, we implemented a modified image-filtering method and developed a restoration model specifically tailored to estimate the shape of fine wire rods. Section 4 presents the details of this approach. 3. Development of the Wire Rod Segmentation Model This section details the development of a wire rod segmentation model designed to distinguish material from background pixels in thermal images. The model was developed by the Furnace Process Research Team at a leading Korean steel manufacturer responsible for quality control in electric furnace-based steel production processes, including steelmaking, rolling, surface treatment, and temperature monitoring. The methodology is described below. 3.1 Data Collection As shown in Fig. 1 , an infrared line scanner installed at the end of the cooling conveyor automatically collected temperature data in the form of two-dimensional thermal images. Wire rods were grouped into seven categories based on diameter, ranging from 5.5 mm to 18 mm. Five groups had diameters greater than 10 mm and were classified as heavy type, while the remaining two groups—with diameters less than 10 mm—were classified as fine type. We collected one thermal image for each group, resulting in five images for heavy types and two for fine types. Figure 3 presents visualizations of these thermal images. Although the original temperature data captured by the scanner were colorless, color mapping was applied for clarity—where red indicates higher temperatures and blue indicates lower temperatures. As shown in Fig. 3 (a), heavy wire rods appear predominantly red, with a distinguishable blue background. For this reason, only the five heavy-type thermal images were used to generate training data, train the model, and validate the results. However, in Fig. 3 (b), the fine-type wire rods are difficult to visually distinguish from the background, making manual segmentation unreliable. Each original thermal image had a horizontal resolution of 1000 pixels and a vertical resolution ranging from thousands to tens of thousands of pixels, depending on the product size. To reduce computational complexity, a 512-pixel-wide segment of temperature data containing the wire rod was extracted from each image for further processing. 3.2 Preprocessing and Annotation Although we reduced the original thermal images to 512 pixels in width, this resolution was still computationally demanding, resulting in high processing costs and memory requirements. Additionally, with only five thermal images available, the dataset was insufficient for effective model training. To address these limitations, we implemented a patch-based approach. Patch- or tile-based methods are commonly used for high-resolution image datasets with limited sample sizes (Yuan et al., 2021 ). Common input dimensions for CNNs include 128×128, 224×224, and 256×256 pixels (Rukundo, 2023 ; Talebi & Milanfar, 2021 ). Each image was divided into 128 × 128-pixel patches with four horizontal segments, enabling separate analysis of central and edge regions, which exhibit different physical properties (Hwang, 2020 ). An even number of horizontal partitions was selected to maintain input consistency and facilitate efficient data augmentation through rotational and reflectional transformations, leveraging the bilateral symmetry of wire rods. The 128×128 patch size was determined to be optimal for preserving structural features while enabling real-time processing. Figure 4 illustrates the image partitioning process. To improve visibility for manual annotation, a color filter was applied to the image patches. The Blue-White-Red (BWR) filter was selected because it intuitively maps high temperatures to red and low temperatures to blue, aiding annotators in distinguishing material from background. After normalization based on the minimum and maximum temperature values within each patch, pixel colors near the maximum temperature appeared red, and those near the minimum appeared blue. Figure 5 demonstrates how the filter application method affects patch visibility. Compared to applying a single filter based on the entire image’s temperature range (Fig. 5 (a)), applying an individual filter to each patch based on its own temperature range (Fig. 5 (b)) exhibits a more pronounced distinction between the material and background, thereby improving the accuracy and consistency of manual annotation. Figure 6 illustrates the key components of a thermal image. Annotators were instructed to identify only the wire rod regions—primarily red zones—while excluding boundary regions, which typically appear white and represent radiant heat dissipating into the surrounding air rather than the material itself. The heavy-type wire rods were segmented into 1292 patches, of which 340 were selected for annotation. These were manually labelled to reflect the visual determination of the wire rod area. The number of selected patches (340) was determined based on the labor required for manual annotation. Figure 7 shows the annotation methodology. Each patch was annotated using two labels: 'wire rod' for the material and 'background' for all non-material areas, including the boundary area. Among the 340 patches, 70% were augmented 8-fold through rotations and flipping. 3.3 Training the Semantic Segmentation Model For pixel-wise classification, we trained three models: the FCN (the most basic model in the field of semantic segmentation), SegNet, and U-Net (an improvement on the FCN). SegNet, which was developed for pixel-wise road scene segmentation, was also employed by Morales-Cervantes et al. ( 2023 ). U-Net, one of the most widely used semantic segmentation models, was adopted and modified for this task. All training sessions used a split of 70% training, 10% validation, and 20% test sets, based on the original 340 patches. Since only the training dataset was augmented, the final ratio of patches across training, validation, and testing was approximately 56:1:2. Figure 8 illustrates the U-Net architecture implemented in this study. The network comprised 22 layers and operated through a sequence of key stages. The input to the model was a 128 × 128 image patch, as generated in Section 3.2 , with a depth of 3 corresponding to red, blue, and green colors. Downsampling was performed through repeated application of 3×3 convolution operations followed by 2×2 max pooling. For the convolution layers, the ReLU activation function was applied, and the same padding was used to preserve spatial dimensions. In contrast to the original U-Net by Ronneberger et al. ( 2015 ), which used no padding and reduced output resolution for computational efficiency, we applied padding to maintain spatial resolution. This modification was necessary because the width of a wire rod in the image can be as small as 1–5 pixels. Without padding, critical structural information may be lost due to excessive downsampling. After downsampling, the feature map was reduced to a resolution of 16 × 16. Upsampling was then performed using 2×2 upconvolutions (transposed convolutions). A skip connection was implemented by concatenating the upsampled feature map with its corresponding feature map from the downsampling path. This skip architecture enabled the model to combine local information from shallow layers with contextual information from deeper layers, enhancing segmentation accuracy. After repeating the final upsampling phase, a 1×1 convolution was applied to produce a semantic segmentation map of the exact resolution as the input (128 × 128). 3.4 Semantic Segmentation Model Results To evaluate the performance of the semantic segmentation models, we used the following quantitative metrics: accuracy, precision, recall, specificity, F1 score, and mIoU. These were calculated based on the confusion matrix. Table 1 presents the confusion matrix index for binary classification. Table 2 lists the formulas used to calculate each evaluation metric. Table 3 summarizes the test performance of each model, both with and without data augmentation. Table 1 Index of confusion matrix. Confusion Matrix Index Predicted Class Wire Rod Background Actual Class Wire Rod TWP (True Wire rod Pixel) FWP (False Background Pixel) Background FWP (False Wire rod Pixel) TBP (True Background Pixel) Table 2 Formulas for quantitative evaluation of segmentation performance. Performance Evaluation Formula Accuracy \(\:\frac{TWP+TBP}{Total\:Samples}\) Precision \(\:\frac{TWP}{TWP+FWP}\) Recall \(\:\frac{TWP}{TWP+FBP}\) Specificity \(\:\frac{TBP}{FWP+TBP}\) F1 Score \(\:\frac{2\bullet\:(Precision\bullet\:Recall)}{Precision+Recall}\) (mIoU) \(\:\frac{Area\:of\:Overlap}{Area\:of\:Union}\) Table 3 Comparing the performance of the models. Original data Augmented data FCN SegNet U-Net FCN SegNet U-Net Accuracy 0.9361 0.9213 0.9502 0.9616 0.9580 0.9624 Precision 0.9662 0.9400 0.9599 0.9725 0.9706 0.9743 Recall 0.9339 0.9381 0.9631 0.9679 0.9642 0.9674 Specificity 0.9403 0.8906 0.9266 0.9500 0.9466 0.9535 F1-score 0.9125 0.8890 0.9294 0.9459 0.9410 0.9473 mIoU 0.8390 0.8003 0.8682 0.8974 0.8885 0.8998 All three models showed improved performance after data augmentation. Notably, SegNet exhibited the most significant improvement, with its mIoU increasing by 0.0971 after augmentation. This improvement is attributed to the bilateral symmetry of wire rod thermal images, which makes rotation- and flipping-based augmentation highly effective for learning. In terms of model comparison, U-Net outperformed both FCN and SegNet across all evaluation metrics. When trained on augmented data, the U-Net achieved an accuracy of 0.9624, an F1 score of 0.9473, and an mIoU of 0.8998. It correctly identified 96.74% (recall) of actual wire rod pixels and 95.35% (specificity) of background pixels, demonstrating strong performance despite the complex morphology of wire rods. In addition to quantitative evaluation, visual validation of model predictions was also performed. This qualitative approach is beneficial for evaluating model generalization on unlabeled data. The trained U-Net model was applied to all 1292 heavy-type wire rod image patches, and the predictions showed high spatial accuracy and consistency in the reconstructed shapes. Even for the 18 mm product—characterized by thicker material and greater thermal variation—the predicted results were deemed acceptable for visual inspection purposes. Figure 9 presents a portion of the predicted image. However, when applied to fine-type wire rod images, which were not included in the training set, the model produced incomplete semantic maps, as shown in Fig. 13 (b). This limitation stems from the significant domain difference between the heavy-type training data and the fine-type inference data. Since supervised models perform poorly on out-of-distribution samples, a dedicated approach is required to address fine-type wire rod segmentation. 4. Development of Fine-type Wire Rod Segmentation Methods Measuring the temperature of only the material in thermal images of fine-type wire rods presents unique challenges. In many cases, even the human eye struggles to distinguish the material from the background due to minimal temperature contrast. Nonetheless, if a model can learn the morphological patterns and curvature typical of wire rods under more distinguishable conditions (for example, heavy-type), estimating material pixels with reasonable accuracy becomes possible. Such estimations are valuable for process monitoring, as they allow for the calculation of temperature statistics—such as the mean and variance—across the production history of a product. This section presents the development of segmentation methods tailored to fine-type wire rods. The proposed solution combines the existing segmentation model (developed in Section 3 ) with a novel interpolation and image-restoration method to create a complex semantic map. 4.1 Sliding Window Patch In thermal images of fine-type wire rods (Fig. 13 (a)), only the outer edges of the wire rod coils—where the material is denser and retains more heat—are shown in red, while the interior regions rapidly cool and appear blue or white. When using the fixed patching method described in Section 3.2 , boundary-related temperature gradients and discontinuities often lead to prediction instability, as shown in Fig. 13 (b). For heavy-type wire rods, the lower temperature at the core of the image patch is still distinguishable from the background due to their thicker cross-sections and slower cooling rates. Conversely, fine-type wire rods cool rapidly, and within a single patch, the material temperature may vary significantly, resembling the background. Consequently, the segmentation model—trained primarily on well-defined high-temperature material—may underestimate or miss portions of the wire rod entirely when applied to fine-type images. To address this issue, we propose a sliding window-based patching method. Instead of dividing the image into four patches, we generated overlapping 128×128 patches by shifting a window every 32 pixels in both horizontal and vertical directions. For each patch, an individual color filter was applied, normalizing the patch independently to capture local contrast better. Algorithm 1 outlines the classification process. Unlike the method described in Section 3.2 , this approach allows up to 16 patches to contribute to the prediction of the same pixel (temperature) at a given location. Algorithm 1: Pixel Classification with Sliding Window Patches # Input: Image I, Segmentation model M, Threshold θ # Output: Final prediction map P for each pixel (i, j) in I do predictions ← empty list for each patch containing pixel (i, j) do pred ← M.predict(patch) append pred[i, j] to predictions end for prediction_sum ← sum of all values in predictions if prediction_sum ≥ θ then P[i, j] ← 1 // Classified as wire rod else P[i, j] ← 0 // Classified as background end if end for return P The overlapping sliding window method is widely utilized in image processing tasks to effectively manage high-resolution images and ensure spatial continuity between adjacent patches (Chen et al., 2019 ). In this study, this approach addresses inconsistencies caused by rigid patch boundaries and enhances segmentation performance on fine-type wire rods. During the prediction phase, where the segmentation model trained in Section 3.3 is applied, a single pixel may be included in multiple overlapping patches, resulting in multiple predictions for that pixel. To resolve conflicts in such overlapping regions, threshold-based fusion techniques are commonly employed at either the window (patch) or pixel level (Fan et al., 2016 ; Guan et al., 2024 ). In our implementation, we adopted a pixel-wise union-based threshold rule: if any prediction classifies a pixel as part of a wire rod, it is ultimately labeled as a wire rod. This deliberate overestimation strategy helps minimize the omission of true material, which is particularly beneficial in low-contrast regions. To aid comprehension, Fig. 10 illustrates the concept of sliding window-based semantic segmentation using 32 × 32 patches (as opposed to the actual 128 × 128), highlighting how overlapping regions contribute to a more stable semantic map. As shown in Fig. 13 (c), applying this method resolved classification gaps, particularly in the central double-ended patch that was previously misclassified in Fig. 13 (b). Moreover, the semantic map exhibits an improved structural consistency, with clearly repeating circular patterns indicative of the wire rod coil structure. Nevertheless, despite overestimating the material areas, residual fragmentation remains, with dotted or broken contours still visible. 4.2 Training the Restoration Model Although the enhanced segmentation map generated by the sliding window method presents a more complete structure, the results often appear visually fragmented to human observers familiar with the continuous curvature of wire rods. These dotted predictions resemble blue noise, introducing ambiguity in determining the true morphology of the wire rods. To enhance the prediction quality, a CNN model can be trained to infer the intact shape of a wire rod from imperfect semantic maps. We refer to this as the restoration model. Figure 11 provides an overview of the restoration process. Training data for this model were derived from heavy-type wire rod semantic maps, which provide more reliable structural patterns. Additionally, we gathered two types of input-output image patch pairs: semantic maps of manually annotated data from Section 3 , and high-quality segments of the predicted semantic maps, as shown in Fig. 10 . Using a sliding window with a stride of 8 pixels, we extracted 3,956 patches of 512 × 512 pixels. Artificial noise was introduced by randomly setting 65% to 95% of the pixels in each input patch to zero, simulating the sparse prediction pattern seen in fine-type wire rod segmentation. The restoration model was trained to reconstruct the original shape from these degraded inputs. We used a standard 7:1:2 ratio to split the dataset into training, validation, and test sets. Figure 12 illustrates the network architecture used for restoration. The architecture of the restoration model is based on U-Net, as shown in Fig. 12 , but differs from the semantic segmentation model (Fig. 8 ) in several key aspects, including input and output dimensions (512 × 512×1), the specific objective of noise suppression, and the morphological restoration of wire rod patterns. 4.3 Results The restoration model, trained to infer the complete morphology of the wire rods from incomplete semantic maps, achieved a test accuracy of 0.9632. This model is designed to reconstruct the full wire rod structure by learning from intentionally degraded input images and their corresponding intact versions, effectively addressing visual noise issues such as the "blue noise" effect observed in Fig. 13 (c). As no labeled data exist for fine-type wire rods, quantitative evaluation is not feasible. Therefore, the semantic map must be assessed quantitatively through visual inspection. Figure 13 (d) shows the results of applying both the sliding window method (Section 4.2 ) and the restoration model (Section 4.3 ) to fine-type wire rods. Compared with the original image in Fig. 13 (a), which is visually ambiguous, the restored semantic map accurately predicts the expected wire rod regions. Notably, Fig. 13 (d) demonstrates a significant improvement over Fig. 13 (c). Although the latter contains noisy, fragmented predictions, the former exhibits smoother, continuous contours, resulting in a more realistic and coherent wire rod shape. This approach leverages the ability of the model to learn and reconstruct structural patterns, even from sparse or noisy inputs, thereby enhancing the overall visual quality and consistency of the wire rod segmentation. However, a limitation remains at the head region of the wire rod, where the prediction accuracy has not improved significantly. Despite this limitation, the current results effectively fulfill the primary objective highlighted in Chap. 4—to facilitate a qualitative understanding of temperature dispersion across the entire wire rod length. 5. Conclusion This study proposes a novel, automated method for acquiring post-cooling temperature data of wire rods using image-processing techniques, specifically semantic segmentation. The proposed methodology is designed to overcome challenges posed by the complex geometry of wire rods and limitations of conventional point-based temperature measurement methods in the steel industry. Our approach consists of three components: (1) a preprocessing technique that divides thermal images into manageable patches and enhances contrast using color filters; (2) a U-Net-based semantic segmentation model trained to differentiate wire rods from the background; and (3) a restoration model that reconstructs fine-type wire rod shapes from noisy predictions. The semantic segmentation model, trained on augmented heavy-wire rod data, achieved an accuracy of 0.9624, an F1-score of 0.9473, and an mIoU of 0.8998—demonstrating its strong performance despite the intricate morphology of wire rods. For fine-type wire rods, we introduced a sliding window-based patch-generation method combined with a restoration model. This hybrid strategy mitigates prediction artifacts, connects fragmented segments, and generates more natural, continuous semantic maps, thereby indicating the applicability of the model to previously untrainable datasets. The significance of this study lies in its ability to automate full-length temperature profiling of wire rods, a task that was previously limited to manual spot-checking. This advancement has direct implications for quality control in wire rod production, with potential benefits including defect reduction and enhanced product uniformity. Future research should focus on acquiring annotated datasets for fine-type wire rods, enhancing the model through domain-specific data augmentation, and integrating the temperature profiling system with predictive models of physical properties, such as hardness prediction models (Pyo et al., 2024 )—to establish a comprehensive quality control framework for wire rod manufacturing. Declarations a. Funding: This work was supported by the Korea Evaluation Institute of Industrial Technology (KEIT) and the Ministry of Trade, Industry, and Energy (MOTIE) of the Republic of Korea (grant number [RS-2022-00155473]: Development of technology for improving energy efficiency and product quality through the application of big data in the steel rolling process). b. Competing Interests: The authors report there are no competing interests to declare. c. Data Availability Statement: Derived data supporting the findings of this study are available from the corresponding author on request. d. Authors' contributions: All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Seok-Kyu Pyo, Dong-Hee Lee, Sung-Jun Hur, Sang-Hyeon Lee, Sung-Jun Lim, Jong-Eun Lee. The first draft of the manuscript was written by Seok-Kyu Pyo and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. References Abraham, G. K. (2013). The Effect of an Encapsulated Specimen’s Mounting Material on Its Rockwell Hardness Test Results. Metallography, Microstructure, and Analysis , 2 (6), 378–382. https://doi.org/10.1007/s13632-013-0104-6 Asthana, N., & Byeon, H. (2024). U-Net-based Recommender Systems for Political Election System using Collaborative Filtering Algorithms. Journal of Information and Communication Convergence Engineering , 22 (1), 7–13. https://doi.org/10.56977/jicce.2024.22.1.7 Che, L., He, Z., Zheng, K., Xu, X., & Zhao, F. (2024). An automatic segmentation and quantification method for austenite and ferrite phases in duplex stainless steel based on deep learning. Journal of Materials Chemistry A . https://doi.org/10.1039/d4ta05421k Chen, M., Zhao, X., & Xu, D. (2019). Image Inpainting for Digital Dunhuang Murals Using Partial Convolutions and Sliding Window Method. Journal of Physics: Conference Series , 1302 (3). https://doi.org/10.1088/1742-6596/1302/3/032040 Daigo, I., Murakami, K., Tajima, K., & Kawakami, R. (2023). Thickness Classifier on Steel in Heavy Melting Scrap by Deep-learning-based Image Analysis. ISIJ International , 63 (1), 197–203. https://doi.org/10.2355/isijinternational.ISIJINT-2022-331 Fan, X., Riaz, I., Rehman, Y., & Shin, H. (2016). Vanishing point detection using random forest and patch-wise weighted soft voting. IET Image Processing , 10 (11), 900–907. https://doi.org/10.1049/iet-ipr.2016.0068 Guan, M., Yu, W., Guo, Y., Huang, K., Zhang, J., Liang, K., & Ma, Z. (2024). Focus the Overlapping Problem on Few-Shot Object Detection via Multiple Predictions. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) , 14426 LNCS , 427–439. https://doi.org/10.1007/978-981-99-8432-9_34 Harste, K., & Wustner, E. (2004). Innovative solutions for the production of wire rod and bar. STAHL UND EISEN , 124 (1), 43–48. Hwang, J. K. (2020). Effect of Ring Configuration on the Deviation in Cooling Rate and Mechanical Properties of a Wire Rod during the Stelmor Cooling Process. Journal of Materials Engineering and Performance , 29 (3), 1732–1740. https://doi.org/10.1007/s11665-020-04694-0 Koymatcik, H., Ahlatci, H., Sun, Y., & Turen, Y. (2018). Effect of carbon content and drawing strain on the fatigue behavior of tire cord filaments. ENGINEERING SCIENCE AND TECHNOLOGY-AN INTERNATIONAL JOURNAL-JESTECH , 21 (3), 289–296. https://doi.org/10.1016/j.jestch.2018.04.010 Laub, M., Bachmann I, B.-, Detemple, E., Scherff, F., Staudt, T., Mueller, M., Britz, D., Muecklich, F., & Motz, C. (2022). Determination of grain size distribution of prior austenite grains through a combination of a modified contrasting method and machine learning. PRAKTISCHE METALLOGRAPHIE-PRACTICAL METALLOGRAPHY , 60 (1), 4–36. https://doi.org/10.1515/pm-2022-1025 Lema, D. G., Pedrayes, O. D., Usamentiaga, R., & García, D. F. (2023). Improved detection of subsurface defects through active thermography and ensembling techniques. Quality Engineering , 0 (0), 1–17. https://doi.org/10.1080/08982112.2023.2177871 Long, J., Shelhamer, E., & Darrell, T. (2015). Fully convolutional networks for semantic segmentation. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 3431–3440. https://doi.org/10.1109/CVPR.2015.7298965 Morales-Cervantes, A., Chávez-Campos, G. M., Vergara-Hernández, H. J., Flores, J. J., & Guevara, E. (2023). Monitoring Steel Heating Processes Using Infrared Thermography and Deep Learning-Based Semantic Segmentation. JOM . https://doi.org/10.1007/s11837-023-06212-8 Pavlina, E. J., & Van Tyne, C. J. (2008). Correlation of Yield Strength and Tensile Strength with Hardness for Steels. JOURNAL OF MATERIALS ENGINEERING AND PERFORMANCE , 17 (6), 888–893. https://doi.org/10.1007/s11665-008-9225-5 Pedrayes, O. D., Lema, D. G., Usamentiaga, R., Venegas, P., & García, D. F. (2022). Semantic segmentation for non-destructive testing with step-heating thermography for composite laminates. Measurement: Journal of the International Measurement Confederation , 200 . https://doi.org/10.1016/j.measurement.2022.111653 Pyo, S. K., Lee, D. H., Hur, S. J., Lee, S. H., Lim, S. J., Lee, J. E., & Moon, H. K. (2024). Prediction of spring steel wire rod hardness based on wire rod rolling process data: a case study. International Journal of Advanced Manufacturing Technology , 133 (1–2), 1011–1024. https://doi.org/10.1007/s00170-024-13740-3 Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) , 9351 , 234–241. https://doi.org/10.1007/978-3-319-24574-4_28 Rukundo, O. (2023). Effects of Image Size on Deep Learning. Electronics (Switzerland) , 12 (4). https://doi.org/10.3390/electronics12040985 Talebi, H., & Milanfar, P. (2021). Learning to Resize Images for Computer Vision Tasks. CoRR , abs/2103.09950 . https://arxiv.org/abs/2103.09950 Thomas, A., Durmaz, A. R., Straub, T., & Eberl, C. (2020). Automated quantitative analyses of fatigue-induced surface damage by deep learning. Materials , 13 (15). https://doi.org/10.3390/ma13153298 Usamentiaga, R., Garcia, D. F., & Perez, J. M. (2020). High-Speed Temperature Monitoring for Steel Strips Using Infrared Line Scanners. IEEE Transactions on Industry Applications , 56 (3), 3261–3271. https://doi.org/10.1109/TIA.2020.2980478 Wei, Y., Li, Y., Zhu, L., Liu, Y., Lei, X., Wang, G., Wu, Y., Mi, Z., Liu, J., Wang, H., & Gao, H. (2014). Evading the strength- ductility trade-off dilemma in steel through gradient hierarchical nanotwins. NATURE COMMUNICATIONS , 5 . https://doi.org/10.1038/ncomms4580 Xie, L., Li, W., Fan, L., & Zhou, M. (2023). Automatic Identification of the Multiphase Microstructures of Steels based on ASPP-FCN. Steel Research International . https://doi.org/10.1002/srin.202200204 Xu, Z.-G., Wang, J., & Wang, L.-Y. (2018). Infrared Image Semantic Segmentation Based on Improved DeepLab and Residual Network. In 2018 10th International Conference on Modelling, Identification and Control (ICMIC) . Yuan, X., Shi, J., & Gu, L. (2021). A review of deep learning methods for semantic segmentation of remote sensing imagery. In Expert Systems with Applications (Vol. 169). Elsevier Ltd. https://doi.org/10.1016/j.eswa.2020.114417 Cite Share Download PDF Status: Published Journal Publication published 12 Mar, 2026 Read the published version in The International Journal of Advanced Manufacturing Technology → Version 1 posted Editorial decision: Major Revisions Needed 23 Nov, 2025 Reviewers agreed at journal 06 Oct, 2025 Reviewers invited by journal 05 Oct, 2025 Editor assigned by journal 18 Sep, 2025 First submitted to journal 15 Sep, 2025 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-7625115","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":524981752,"identity":"b16ccb4d-49fd-48a0-9127-becc02c9831b","order_by":0,"name":"Seok-Kyu Pyo","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Seok-Kyu","middleName":"","lastName":"Pyo","suffix":""},{"id":524981753,"identity":"cb66595d-450f-467e-8975-cc489099b47e","order_by":1,"name":"Dong-Hee Lee","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAoUlEQVRIiWNgGAWjYDCCG0D8gSHBAMSWIFoL4wyStTDzkKSF73bzwcc2f9KMDQ4wH7zNQ4wWyTvHko1z23LMDA6wJVsTpcXgRo6ZdG5DhY3BAR4zaSK15H//bfEHpIX/G7FactiYGdhADuNhI06L5I00Y8neNiBxmM3Ycg4xWvhuJD/88ONPsmHf8eaHN94QowUBmElTPgpGwSgYBaMAHwAA9UMx8FfR1iYAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0001-8549-8992","institution":"Sungkyunkwan University","correspondingAuthor":true,"prefix":"","firstName":"Dong-Hee","middleName":"","lastName":"Lee","suffix":""},{"id":524981754,"identity":"26aa864f-8b4c-45fd-9147-f8d2748e5a0e","order_by":2,"name":"Sung-Jun Hur","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Sung-Jun","middleName":"","lastName":"Hur","suffix":""},{"id":524981755,"identity":"4a2af377-c831-43ab-a029-074c6cf930d8","order_by":3,"name":"Sang-Hyeon Lee","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Sang-Hyeon","middleName":"","lastName":"Lee","suffix":""},{"id":524981756,"identity":"f25d6858-3e0f-49a7-b1f3-973bfc537e40","order_by":4,"name":"Sung-Jun Lim","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Sung-Jun","middleName":"","lastName":"Lim","suffix":""},{"id":524981757,"identity":"4124f60a-592f-4e1a-b711-a38014a497d0","order_by":5,"name":"Jeong-Eun Lee","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Jeong-Eun","middleName":"","lastName":"Lee","suffix":""}],"badges":[],"createdAt":"2025-09-16 02:49:03","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7625115/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7625115/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s00170-026-17717-2","type":"published","date":"2026-03-12T15:57:50+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":93731834,"identity":"fa116a05-1911-40e9-b0d2-45c5a94ce6bc","added_by":"auto","created_at":"2025-10-17 02:24:12","extension":"xml","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8817,"visible":true,"origin":"","legend":"","description":"","filename":"jamtJAMTD2505332.xml","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/e527359ac4423fe52e8c5a6a.xml"},{"id":93731849,"identity":"dd3aa558-7aa1-41af-a6b6-fb35cb16bbb8","added_by":"auto","created_at":"2025-10-17 02:24:13","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":936,"visible":true,"origin":"","legend":"","description":"","filename":"JAMTD2505332156284.go.xml","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/289c59233a423cb41e76a8a7.xml"},{"id":93731858,"identity":"8b0b77f5-4462-433f-8c90-05f971eb8861","added_by":"auto","created_at":"2025-10-17 02:24:14","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":824,"visible":true,"origin":"","legend":"","description":"","filename":"JAMTD2505332Import.xml","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/f9d1e79477fde9dc63343833.xml"},{"id":93731827,"identity":"1edfca58-715b-444f-b69d-99b9704cb6dc","added_by":"auto","created_at":"2025-10-17 02:24:11","extension":"xml","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":97804,"visible":true,"origin":"","legend":"","description":"","filename":"JAMTD25053320enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/c7b9ac1285d3a8358636e7cf.xml"},{"id":93731861,"identity":"8fa4bafc-368e-44aa-acf2-37afb93566d8","added_by":"auto","created_at":"2025-10-17 02:24:15","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":483742,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/0396b35fc7ddf31fcf204876.png"},{"id":93731860,"identity":"828de439-08f9-4b31-b38d-bee7d0be8708","added_by":"auto","created_at":"2025-10-17 02:24:14","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":840158,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/6d81b725733d8325d56e41b9.png"},{"id":93731882,"identity":"3f4ff234-4c98-4fe2-be23-bc6b8651d75f","added_by":"auto","created_at":"2025-10-17 02:24:18","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":180399,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/93a155083a81190089a1ebb1.png"},{"id":93731836,"identity":"6ee6743e-d1e7-46f8-9d7d-e3bd4f87c392","added_by":"auto","created_at":"2025-10-17 02:24:12","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":174222,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/43880de7d8c96d8bc317515d.png"},{"id":93732535,"identity":"cd5e55a0-f933-4150-a888-5eb79255266c","added_by":"auto","created_at":"2025-10-17 02:32:15","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1782828,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/6c80f14db68338bae56ffedf.png"},{"id":93731826,"identity":"cd871b9e-d97a-4540-ab0e-4a9302cbb0b4","added_by":"auto","created_at":"2025-10-17 02:24:11","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":260110,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/5c54404602f60846100c6250.png"},{"id":93731869,"identity":"8c90e4b6-464c-4ec9-9e4d-d0884758bdf2","added_by":"auto","created_at":"2025-10-17 02:24:15","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":696709,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/bf58f5cb16d4087e2aee3057.png"},{"id":93731865,"identity":"41c30812-ffee-4d4c-909a-854aef28d12e","added_by":"auto","created_at":"2025-10-17 02:24:15","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":384722,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/e1031b5f7c70bcf832549873.png"},{"id":93731822,"identity":"55c05e3e-fb58-4a8a-9c02-2461289edf8a","added_by":"auto","created_at":"2025-10-17 02:24:11","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":604937,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/51b1f575c429e206848ca636.png"},{"id":93731881,"identity":"49af6b51-abb4-4758-8a25-162501f9a26e","added_by":"auto","created_at":"2025-10-17 02:24:18","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":534758,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/88c909c53febc7f8f1d489c3.png"},{"id":93731846,"identity":"8d5b2e31-d995-4933-9c5a-fa18fc90d44a","added_by":"auto","created_at":"2025-10-17 02:24:13","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":651269,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/2a75f4faffb3e2b70823a505.png"},{"id":93731884,"identity":"302ed849-3b95-448a-8805-d11bd9495eac","added_by":"auto","created_at":"2025-10-17 02:24:19","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":185653,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/e43b60e7e6ad19684862ce84.png"},{"id":93731864,"identity":"03da7f20-fbea-48bd-82ea-76a3ccacd1b6","added_by":"auto","created_at":"2025-10-17 02:24:15","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":952872,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/d09973ad5cd945f7c25d1a76.png"},{"id":93731840,"identity":"abc9a643-cf95-4f2d-aeaf-f22f3b9713f1","added_by":"auto","created_at":"2025-10-17 02:24:12","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":71010,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/6d0d05026c13816b4fe19410.png"},{"id":93732534,"identity":"542e01b4-120f-4b42-854a-890163c5ed34","added_by":"auto","created_at":"2025-10-17 02:32:13","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":87662,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/88eb8a018027c5d4df850b9f.png"},{"id":93731851,"identity":"faee8bc7-59cd-4ce1-8ccc-7431dbe2a9e3","added_by":"auto","created_at":"2025-10-17 02:24:13","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61028,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/dbe2965cee4802e9bb162649.png"},{"id":93731874,"identity":"17d110ab-f474-4667-a162-5363be87fbbd","added_by":"auto","created_at":"2025-10-17 02:24:16","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":60514,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/52c0d62640522372938e4f92.png"},{"id":93732530,"identity":"ea5eedb7-c49d-47b8-9904-b3704e5ce255","added_by":"auto","created_at":"2025-10-17 02:32:11","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":350590,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/918bbbcc0625b524810c76ce.png"},{"id":93731852,"identity":"636283d9-b21a-485e-bc9d-0dbd61a39519","added_by":"auto","created_at":"2025-10-17 02:24:13","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":31558,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/a5689b3a16d5c603b6e91f33.png"},{"id":93731871,"identity":"4a1ec90c-28f1-46ef-adba-500b93a09e3f","added_by":"auto","created_at":"2025-10-17 02:24:15","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":118143,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/1b32761b90603dff5972010a.png"},{"id":93732529,"identity":"35f5e9dc-a388-409e-aebb-543ad4d25e86","added_by":"auto","created_at":"2025-10-17 02:32:11","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":67193,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/2e894ad407dd2e96afbd0c27.png"},{"id":93731835,"identity":"3f0b4547-7d5c-4d29-be59-7aa581c9f689","added_by":"auto","created_at":"2025-10-17 02:24:12","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":74416,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/f5a8a32be7fb2f1c133afbc4.png"},{"id":93731820,"identity":"b4634e8e-b44b-47c4-9087-e2ff22677e7c","added_by":"auto","created_at":"2025-10-17 02:24:11","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":46605,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/a9cfbcbf46ba28b18b5c4ebf.png"},{"id":93731859,"identity":"5321d995-b99d-4a5d-86c0-4b1c353280cc","added_by":"auto","created_at":"2025-10-17 02:24:14","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":66292,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/c6701b06cddfc73b37809f4b.png"},{"id":93731848,"identity":"9de27e92-0905-4110-afc4-c34347f45e3c","added_by":"auto","created_at":"2025-10-17 02:24:13","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":44865,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/97d6cbcc490ad35767d62545.png"},{"id":93731839,"identity":"eb48099b-fbc1-4609-88df-61159a239e33","added_by":"auto","created_at":"2025-10-17 02:24:12","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":316818,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/64894e9f0003153c9308ef3b.png"},{"id":93731857,"identity":"6986eb19-991f-40e5-9a92-205464e9c0ed","added_by":"auto","created_at":"2025-10-17 02:24:14","extension":"xml","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":96948,"visible":true,"origin":"","legend":"","description":"","filename":"JAMTD25053320structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/85197d305c02997f2d6d7901.xml"},{"id":93731888,"identity":"1ade5fc4-a2ad-4bed-b151-364e12aded86","added_by":"auto","created_at":"2025-10-17 02:24:19","extension":"html","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":104722,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/8d7f4e8ffbf04e9c90482614.html"},{"id":93731824,"identity":"53a13ebe-b850-41b2-9c82-ca3c8871c787","added_by":"auto","created_at":"2025-10-17 02:24:11","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":633381,"visible":true,"origin":"","legend":"\u003cp\u003eWire rod thermal image and components.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/2d9f42cfc0ef6c13ff6892d4.png"},{"id":93732533,"identity":"274be358-9f1a-40e1-8e1f-c9a0560d426f","added_by":"auto","created_at":"2025-10-17 02:32:12","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":224958,"visible":true,"origin":"","legend":"\u003cp\u003eDifferences between product types in collecting material temperature from thermal images.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/872e0e775de64c7af573c5db.png"},{"id":93731821,"identity":"d0d610ee-3f25-462c-8bb9-a9e608fcf95d","added_by":"auto","created_at":"2025-10-17 02:24:11","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":426176,"visible":true,"origin":"","legend":"\u003cp\u003eThermal images of (a) heavy-type wire rod and (b) fine-type wire rod\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/6c48086b132df1051857315a.png"},{"id":93731830,"identity":"589c20e9-ca9d-43b3-a0a8-7603e4ceaf78","added_by":"auto","created_at":"2025-10-17 02:24:11","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":275262,"visible":true,"origin":"","legend":"\u003cp\u003eDivision of a heavy-type wire rod thermal image into 128×128 patches\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/638329d96c07776a67261dab.png"},{"id":93731814,"identity":"165bf531-0277-4749-9ecf-67d5ada65554","added_by":"auto","created_at":"2025-10-17 02:24:09","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":478099,"visible":true,"origin":"","legend":"\u003cp\u003eEffect of color filter application method on image patches: (a) a single filter applied globally to the entire image; (b) an individual filter applied per patch.\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/ee2322242c11db4be932fafb.png"},{"id":93732531,"identity":"007f05f9-a5ca-418a-9916-73e76b7ea3dc","added_by":"auto","created_at":"2025-10-17 02:32:12","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":348501,"visible":true,"origin":"","legend":"\u003cp\u003eThermal image components of a wire rod\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/40d307cce422833f42f5bb3f.png"},{"id":93732532,"identity":"69084561-0207-4873-8676-fc87dd2f8d3d","added_by":"auto","created_at":"2025-10-17 02:32:12","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":514182,"visible":true,"origin":"","legend":"\u003cp\u003eManual annotation of a wire rod thermal image.\u003c/p\u003e","description":"","filename":"image7.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/1e9199966215047c75a214f6.png"},{"id":93731845,"identity":"bb6a36c1-fc7f-4807-aede-2af0245a122f","added_by":"auto","created_at":"2025-10-17 02:24:13","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":344588,"visible":true,"origin":"","legend":"\u003cp\u003eU-Net architecture implemented for wire rod thermal image segmentation.\u003c/p\u003e","description":"","filename":"image8.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/0011dfa6d9ac70aae5dc7fee.png"},{"id":93731838,"identity":"8e13ad77-3851-482f-b1e4-330b2eae51ae","added_by":"auto","created_at":"2025-10-17 02:24:12","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":1831035,"visible":true,"origin":"","legend":"\u003cp\u003eWire rod prediction images.\u003c/p\u003e","description":"","filename":"image9.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/64c8553b63af9ba9027bf275.png"},{"id":93731823,"identity":"6a05261d-a5ca-4530-84ec-520f42ddecf0","added_by":"auto","created_at":"2025-10-17 02:24:11","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":628448,"visible":true,"origin":"","legend":"\u003cp\u003eExample of wire rod semantic segmentation using sliding window patches.\u003c/p\u003e","description":"","filename":"image10.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/8adeb1092c908d7626c48adb.png"},{"id":93731833,"identity":"ecb8c7de-1705-4515-b285-c03c354d3d1f","added_by":"auto","created_at":"2025-10-17 02:24:12","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":283818,"visible":true,"origin":"","legend":"\u003cp\u003eNoising and restoring the semantic map.\u003c/p\u003e","description":"","filename":"image11.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/06d7b5ac2b8ddb5b3ee93ebc.png"},{"id":93731842,"identity":"137082d4-d9e2-466b-ac55-b528e744196b","added_by":"auto","created_at":"2025-10-17 02:24:12","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":394810,"visible":true,"origin":"","legend":"\u003cp\u003eU-Net for restoring the wire rod semantic map.\u003c/p\u003e","description":"","filename":"image12.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/2a3a7c5f86a419f020972c34.png"},{"id":93731829,"identity":"9096675a-e341-469d-9a53-1cd5dc335f64","added_by":"auto","created_at":"2025-10-17 02:24:11","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":1903388,"visible":true,"origin":"","legend":"\u003cp\u003e(a) The original temperature image of the fine-type wire rod, (b) the semantic map predicted by the semantic segmentation model, (c) the semantic map with the sliding window patch and the semantic segmentation model, (d) the semantic map with the previous methodology and the restoration model.\u003c/p\u003e","description":"","filename":"image13.png","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/a38e9bf10c2e447785da3961.png"},{"id":104739304,"identity":"d0f6a4f1-50ae-4d90-8041-573cca24a990","added_by":"auto","created_at":"2026-03-16 16:01:11","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":8944250,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7625115/v1/6df3f4fe-283b-41a1-92ae-316651133fae.pdf"}],"financialInterests":"","formattedTitle":"Semantic Segmentation of Steel Wire Rod Thermal Images for Automated Temperature Measurement","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eA wire rod is a semi-finished steel product that must meet specific requirements regarding its physical properties. Wire rods delivered to customers are further processed into finished products such as wires, springs, bolts, and nuts. Noncompliance with these physical specifications may result in product failure or a reduction in the quality of the finished product. Key physical properties include fatigue strength, tensile strength, yield strength, and hardness (Harste \u0026amp; Wustner, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2004\u003c/span\u003e). Among these properties, hardness is the most important, particularly for automotive suspension springs. When the hardness does not meet specifications, fatigue strength and ductility typically decrease (Koymatcik et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Pavlina \u0026amp; Van Tyne, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2008\u003c/span\u003e; Wei et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Reduced ductility can lead to damage during spring drawing, while decreased fatigue strength compromises the durability and safety of the suspension system. In extreme cases, product fractures can result in costly rework and reputational damage for wire rod manufacturers.\u003c/p\u003e\u003cp\u003eTo improve the quality of wire rods, it is necessary to both measure and predict their hardness. Rockwell hardness testing is a conventional measurement method based on specimen sampling (Abraham, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). However, because wire rods are coiled into long lengths, samples are typically taken only from the ends, limiting the comprehensiveness of the inspection. Furthermore, wire rods are hundreds to thousands of meters long, and hardness can vary significantly depending on the position within the coil (Hwang, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Even after inspection, defects may still exist in uninspected sections. This limitation can be mitigated by leveraging historical process data to estimate hardness across the entire wire rod (i.e., full-length hardness).\u003c/p\u003e\u003cp\u003eOne promising approach for estimating full-length hardness is to identify and model the critical factors that influence it. Among the wire rod production stages, this study focuses on the cooling process, which plays a central role in determining hardness. The final transformation occurs during the cooling phase. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates a schematic of the cooling process. In this stage, straight wire rods are cooled to the target temperature using water-cooling boxes and are then coiled into loops via the laying head onto the cooling conveyor. These overlapping loops are cooled with forced air using fans, allowing engineers to achieve the desired microstructure and mechanical properties. The cooling rate is a critical factor affecting hardness (Hwang, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Therefore, the cooling rate, along with pre-cooling and post-cooling temperatures, must be measured.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eMeasuring the post-cooling temperature of a wire rod is considerably more challenging than measuring the pre-cooling temperature. In the steel-rolling process, temperature is typically measured using an infrared line scanner (Usamentiaga et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). The thermal image obtained from the line scanner contains pixel-by-pixel temperature values. However, these thermal images do not clearly differentiate between pixels representing the material surface (that is, the wire rod) and those representing the background. Since the background pixels are noisy and can distort temperature readings, identifying the material pixels is essential. The complexity of this task depends on the geometry of the material. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e illustrates the differences between thermal images of a billet and a wire rod. Billets and steel plates have relatively simple shapes, making it easier to distinguish material regions from the background on a region-by-region basis. Conversely, wire rods processed into coils on a laying head exhibit complex geometries, making it challenging to distinguish material from background. Therefore, engineers often manually collect the temperature values from only a few pixels. This manual approach is time-consuming, yields limited data, and increases the likelihood of human error, leading to higher quality control costs.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eSemantic segmentation offers a promising solution to this challenge. This technique classifies each pixel in an image based on its associated object or region, allowing for the differentiation of material from background across arbitrarily sized image spaces (Long et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). In this context, identifying wire rod pixels in thermal images can be framed as a binary segmentation problem: classifying pixels as either material or background. A semantic segmentation model trained using a convolutional neural network (CNN) architecture can automatically perform this classification. Semantic segmentation techniques have been applied to general photography and electron microscopy images to distinguish between different materials and microstructures (Che et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). However, to the best of our knowledge, no prior studies have applied segmentation techniques to thermal images of wire rods due to their complex shapes.\u003c/p\u003e\u003cp\u003eThis study aims to develop a segmentation model capable of distinguishing material pixels from background pixels in thermal images of wire rods. The model development process includes the collection of thermal images, image processing, augmentation, semantic segmentation, model training, and validation. Each collected thermal image was divided into patches to enable effective learning, reflecting the fact that a single wire rod spans a very long length during rolling. Individual filtering criteria were applied to each patch to ensure a clear distinction between material and background. Manual annotation was used to generate the training data. To enhance training performance, the data were augmented by applying rotations and inversions, taking advantage of the overall vertical and horizontal symmetry in each patch.\u003c/p\u003e\u003cp\u003eWire rods can be categorized into two types: heavy and fine, based on their diameters. Identifying material pixels in fine-type rods is more challenging, even visually, due to the reduced temperature contrast between the material and background caused by rapid cooling. To address this issue, we propose a restoration technique based on a sliding window approach, which enhances the segmentation quality for fine wire rods.\u003c/p\u003e\u003cp\u003eThe remainder of this paper is organized as follows. Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e reviews the relevant literature. Section \u003cspan refid=\"Sec3\" class=\"InternalRef\"\u003e3\u003c/span\u003e describes the proposed segmentation approach. Section \u003cspan refid=\"Sec8\" class=\"InternalRef\"\u003e4\u003c/span\u003e explains the additional procedures implemented to improve segmentation performance for fine wire rods. Section \u003cspan refid=\"Sec12\" class=\"InternalRef\"\u003e5\u003c/span\u003e summarizes the proposed method and discusses directions for future research.\u003c/p\u003e"},{"header":"2. Literature review","content":"\u003cp\u003eThis section reviews semantic segmentation methods and their applications in the steel industry and other manufacturing domains. Semantic segmentation is a central research area in computer vision that involves classifying each pixel in an image into a specific category. Unlike image classification, which assigns a label to an entire image, or object detection, which identifies and locates discrete objects within an image, semantic segmentation provides detailed, pixel-level classification. A foundational model in this field is the Fully Convolutional Network (FCN), which laid the groundwork for subsequent models such as U-Net and DeepLabV3+ (Yuan et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). U-Net is the most cited model in semantic segmentation, with over 30,000 citations of its original paper (Asthana \u0026amp; Byeon, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The name \"U-Net\" derives from its symmetric encoder-decoder architecture, which resembles the letter \u0026ldquo;U.\u0026rdquo; The original U-Net model was designed for biomedical image segmentation, specifically for identifying cells (Ronneberger et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2015\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eIn the steel industry, semantic segmentation has been primarily utilized for classifying steel scrap, analyzing steel microstructure, and detecting defects during heat treatment. For example, Daigo et al. (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) applied semantic segmentation to images of steel scrap to classify different types without direct thickness or diameter measurements. Since different scrap types exhibit varying scales and textures, FCNs often perform poorly due to their limited resolution in deeper layers, which is significantly lower than their resolution in earlier layers. This limitation makes it difficult to detect both fine and coarse features. To address this issue, they proposed a Pyramid Scene Parsing Network (PSPNet), which integrates multi-scale feature representations through a pyramid pooling module.\u003c/p\u003e\u003cp\u003eLaub et al. (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2022\u003c/span\u003e),Thomas et al. (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), andXie et al. (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) applied semantic segmentation to classify steel microstructures. BothLaub et al. (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) andThomas et al. (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) employed U-Net models and performed data augmentation techniques such as rotation and inversion.Laub et al. (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) further selected DenseNet as the encoder backbone through empirical evaluation.Xie et al. (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) proposed an enhanced Atrous Spatial Pyramid Pooling-FCN to identify multiphase microstructures in steel automatically, and conducted comparative experiments across several networks, including FCN, DeepLab v3+, U-Net, Enet, and PSPNet.\u003c/p\u003e\u003cp\u003eMorales-Cervantes et al. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) applied semantic segmentation using CNNs to thermograms for detecting oxidation and decarburization defects during steel heat treatment. Infrared images were obtained during the heating and cooling of a cylindrical steel sample measuring 26 mm in length. The researchers manually labeled each pixel of the images as either background or sample and trained a SegNet model for segmentation. This model achieved a mean accuracy of 92.35% and a mean Intersection over Union (mIoU) of 85.77%. This study shares similarities with our work in that both studies segment thermographic images of steel to distinguish background from material regions. However, our approach differs in several key aspects: we address wire rods that are large, thin, and geometrically complex, and our pipeline includes additional tasks such as image patch segmentation, color filtering, and a restoration model to enhance segmentation quality.\u003c/p\u003e\u003cp\u003eSemantic segmentation of thermal images has also been explored in other manufacturing sectors. Xu et al. (\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) applied DeepLab to segment infrared thermographic images of aluminum electrolysis cells to distinguish between the electrolyte and surface floaters\u0026mdash;an important step for accurate temperature measurement. The irregular shapes and inconsistent distribution of the electrolyte posed significant segmentation challenges. Similarly, Lema et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), Pedrayes et al. (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) applied semantic segmentation to thermography as a non-destructive technique for detecting subsurface defects in carbon fibers. They employed post-processing methods, such as principal component tomography (PCT), in conjunction with U-Net and DeepLabV3\u0026thinsp;+\u0026thinsp;models to enhance segmentation performance.\u003c/p\u003e\u003cp\u003eOverall, our review indicates that semantic segmentation can be successfully applied to thermal images using various deep learning architectures such as FCN, U-Net, and DeepLabV3+. Various preprocessing strategies, including patch-based segmentation and filtering, as well as postprocessing techniques like PCT, have been used to improve model performance. However, no prior studies have applied segmentation techniques to thermal images of wire rods\u0026mdash;a domain characterized by complex geometries, irregular placements, and temperature distributions that critically impact product quality. Therefore, applying semantic segmentation to this context is both novel and appropriate.\u003c/p\u003e\u003cp\u003eAdditionally, existing research has not adequately addressed scenarios in which the temperature contrast between the wire rod and background is minimal, such as in fine-type wire rods. This low contrast presents challenges for conventional preprocessing methods. To resolve this issue, we implemented a modified image-filtering method and developed a restoration model specifically tailored to estimate the shape of fine wire rods. Section \u003cspan refid=\"Sec8\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents the details of this approach.\u003c/p\u003e"},{"header":"3. Development of the Wire Rod Segmentation Model","content":"\u003cp\u003eThis section details the development of a wire rod segmentation model designed to distinguish material from background pixels in thermal images. The model was developed by the Furnace Process Research Team at a leading Korean steel manufacturer responsible for quality control in electric furnace-based steel production processes, including steelmaking, rolling, surface treatment, and temperature monitoring. The methodology is described below.\u003c/p\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Data Collection\u003c/h2\u003e\u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, an infrared line scanner installed at the end of the cooling conveyor automatically collected temperature data in the form of two-dimensional thermal images. Wire rods were grouped into seven categories based on diameter, ranging from 5.5 mm to 18 mm. Five groups had diameters greater than 10 mm and were classified as heavy type, while the remaining two groups\u0026mdash;with diameters less than 10 mm\u0026mdash;were classified as fine type. We collected one thermal image for each group, resulting in five images for heavy types and two for fine types.\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents visualizations of these thermal images. Although the original temperature data captured by the scanner were colorless, color mapping was applied for clarity\u0026mdash;where red indicates higher temperatures and blue indicates lower temperatures. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(a), heavy wire rods appear predominantly red, with a distinguishable blue background. For this reason, only the five heavy-type thermal images were used to generate training data, train the model, and validate the results. However, in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e(b), the fine-type wire rods are difficult to visually distinguish from the background, making manual segmentation unreliable. Each original thermal image had a horizontal resolution of 1000 pixels and a vertical resolution ranging from thousands to tens of thousands of pixels, depending on the product size. To reduce computational complexity, a 512-pixel-wide segment of temperature data containing the wire rod was extracted from each image for further processing.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e3.2 Preprocessing and Annotation\u003c/h2\u003e\u003cp\u003eAlthough we reduced the original thermal images to 512 pixels in width, this resolution was still computationally demanding, resulting in high processing costs and memory requirements. Additionally, with only five thermal images available, the dataset was insufficient for effective model training. To address these limitations, we implemented a patch-based approach. Patch- or tile-based methods are commonly used for high-resolution image datasets with limited sample sizes (Yuan et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Common input dimensions for CNNs include 128\u0026times;128, 224\u0026times;224, and 256\u0026times;256 pixels (Rukundo, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Talebi \u0026amp; Milanfar, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Each image was divided into 128 \u0026times; 128-pixel patches with four horizontal segments, enabling separate analysis of central and edge regions, which exhibit different physical properties (Hwang, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). An even number of horizontal partitions was selected to maintain input consistency and facilitate efficient data augmentation through rotational and reflectional transformations, leveraging the bilateral symmetry of wire rods. The 128\u0026times;128 patch size was determined to be optimal for preserving structural features while enabling real-time processing. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e illustrates the image partitioning process.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo improve visibility for manual annotation, a color filter was applied to the image patches. The Blue-White-Red (BWR) filter was selected because it intuitively maps high temperatures to red and low temperatures to blue, aiding annotators in distinguishing material from background. After normalization based on the minimum and maximum temperature values within each patch, pixel colors near the maximum temperature appeared red, and those near the minimum appeared blue. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e demonstrates how the filter application method affects patch visibility. Compared to applying a single filter based on the entire image\u0026rsquo;s temperature range (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a)), applying an individual filter to each patch based on its own temperature range (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(b)) exhibits a more pronounced distinction between the material and background, thereby improving the accuracy and consistency of manual annotation.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e illustrates the key components of a thermal image. Annotators were instructed to identify only the wire rod regions\u0026mdash;primarily red zones\u0026mdash;while excluding boundary regions, which typically appear white and represent radiant heat dissipating into the surrounding air rather than the material itself.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe heavy-type wire rods were segmented into 1292 patches, of which 340 were selected for annotation. These were manually labelled to reflect the visual determination of the wire rod area. The number of selected patches (340) was determined based on the labor required for manual annotation. Figure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e shows the annotation methodology. Each patch was annotated using two labels: 'wire rod' for the material and 'background' for all non-material areas, including the boundary area. Among the 340 patches, 70% were augmented 8-fold through rotations and flipping.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e3.3 Training the Semantic Segmentation Model\u003c/h2\u003e\u003cp\u003eFor pixel-wise classification, we trained three models: the FCN (the most basic model in the field of semantic segmentation), SegNet, and U-Net (an improvement on the FCN). SegNet, which was developed for pixel-wise road scene segmentation, was also employed by Morales-Cervantes et al. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). U-Net, one of the most widely used semantic segmentation models, was adopted and modified for this task. All training sessions used a split of 70% training, 10% validation, and 20% test sets, based on the original 340 patches. Since only the training dataset was augmented, the final ratio of patches across training, validation, and testing was approximately 56:1:2.\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e illustrates the U-Net architecture implemented in this study. The network comprised 22 layers and operated through a sequence of key stages. The input to the model was a 128 \u0026times; 128 image patch, as generated in Section \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e3.2\u003c/span\u003e, with a depth of 3 corresponding to red, blue, and green colors. Downsampling was performed through repeated application of 3\u0026times;3 convolution operations followed by 2\u0026times;2 max pooling. For the convolution layers, the ReLU activation function was applied, and the same padding was used to preserve spatial dimensions. In contrast to the original U-Net by Ronneberger et al. (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2015\u003c/span\u003e), which used no padding and reduced output resolution for computational efficiency, we applied padding to maintain spatial resolution. This modification was necessary because the width of a wire rod in the image can be as small as 1\u0026ndash;5 pixels. Without padding, critical structural information may be lost due to excessive downsampling. After downsampling, the feature map was reduced to a resolution of 16 \u0026times; 16. Upsampling was then performed using 2\u0026times;2 upconvolutions (transposed convolutions). A skip connection was implemented by concatenating the upsampled feature map with its corresponding feature map from the downsampling path. This skip architecture enabled the model to combine local information from shallow layers with contextual information from deeper layers, enhancing segmentation accuracy. After repeating the final upsampling phase, a 1\u0026times;1 convolution was applied to produce a semantic segmentation map of the exact resolution as the input (128 \u0026times; 128).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e3.4 Semantic Segmentation Model Results\u003c/h2\u003e\u003cp\u003eTo evaluate the performance of the semantic segmentation models, we used the following quantitative metrics: accuracy, precision, recall, specificity, F1 score, and mIoU. These were calculated based on the confusion matrix. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents the confusion matrix index for binary classification. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e lists the formulas used to calculate each evaluation metric. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e summarizes the test performance of each model, both with and without data augmentation.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eIndex of confusion matrix.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" morerows=\"1\" nameend=\"c2\" namest=\"c1\" rowspan=\"2\"\u003e\u003cp\u003eConfusion Matrix Index\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e\u003cp\u003ePredicted Class\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eWire Rod\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eBackground\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eActual Class\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eWire Rod\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTWP (True Wire rod Pixel)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eFWP (False Background Pixel)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eBackground\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eFWP (False Wire rod Pixel)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTBP (True Background Pixel)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eFormulas for quantitative evaluation of segmentation performance.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePerformance Evaluation\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eFormula\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{TWP+TBP}{Total\\:Samples}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{TWP}{TWP+FWP}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{TWP}{TWP+FBP}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{TBP}{FWP+TBP}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eF1 Score\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{2\\bullet\\:(Precision\\bullet\\:Recall)}{Precision+Recall}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e(mIoU)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{Area\\:of\\:Overlap}{Area\\:of\\:Union}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparing the performance of the models.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"9\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colspan=\"4\" nameend=\"c6\" namest=\"c3\"\u003e\u003cp\u003eOriginal data\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003eAugmented data\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003eFCN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSegNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eU-Net\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003eFCN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003eSegNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003eU-Net\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003e0.9361\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.9213\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.9502\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e0.9616\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.9580\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.9624\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003e0.9662\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.9400\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.9599\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e0.9725\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.9706\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.9743\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003e0.9339\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.9381\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.9631\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e0.9679\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.9642\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.9674\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003e0.9403\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.8906\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.9266\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e0.9500\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.9466\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.9535\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eF1-score\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003e0.9125\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.8890\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.9294\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e0.9459\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.9410\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.9473\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003emIoU\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003e0.8390\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.8003\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.8682\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e0.8974\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.8885\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.8998\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eAll three models showed improved performance after data augmentation. Notably, SegNet exhibited the most significant improvement, with its mIoU increasing by 0.0971 after augmentation. This improvement is attributed to the bilateral symmetry of wire rod thermal images, which makes rotation- and flipping-based augmentation highly effective for learning. In terms of model comparison, U-Net outperformed both FCN and SegNet across all evaluation metrics. When trained on augmented data, the U-Net achieved an accuracy of 0.9624, an F1 score of 0.9473, and an mIoU of 0.8998. It correctly identified 96.74% (recall) of actual wire rod pixels and 95.35% (specificity) of background pixels, demonstrating strong performance despite the complex morphology of wire rods.\u003c/p\u003e\u003cp\u003eIn addition to quantitative evaluation, visual validation of model predictions was also performed. This qualitative approach is beneficial for evaluating model generalization on unlabeled data. The trained U-Net model was applied to all 1292 heavy-type wire rod image patches, and the predictions showed high spatial accuracy and consistency in the reconstructed shapes. Even for the 18 mm product\u0026mdash;characterized by thicker material and greater thermal variation\u0026mdash;the predicted results were deemed acceptable for visual inspection purposes. Figure\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e presents a portion of the predicted image. However, when applied to fine-type wire rod images, which were not included in the training set, the model produced incomplete semantic maps, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(b). This limitation stems from the significant domain difference between the heavy-type training data and the fine-type inference data. Since supervised models perform poorly on out-of-distribution samples, a dedicated approach is required to address fine-type wire rod segmentation.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Development of Fine-type Wire Rod Segmentation Methods","content":"\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eMeasuring the temperature of only the material in thermal images of fine-type wire rods presents unique challenges. In many cases, even the human eye struggles to distinguish the material from the background due to minimal temperature contrast. Nonetheless, if a model can learn the morphological patterns and curvature typical of wire rods under more distinguishable conditions (for example, heavy-type), estimating material pixels with reasonable accuracy becomes possible. Such estimations are valuable for process monitoring, as they allow for the calculation of temperature statistics\u0026mdash;such as the mean and variance\u0026mdash;across the production history of a product.\u003c/p\u003e\u003cp\u003eThis section presents the development of segmentation methods tailored to fine-type wire rods. The proposed solution combines the existing segmentation model (developed in Section \u003cspan refid=\"Sec3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) with a novel interpolation and image-restoration method to create a complex semantic map.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Sliding Window Patch\u003c/h2\u003e\u003cp\u003eIn thermal images of fine-type wire rods (Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e (a)), only the outer edges of the wire rod coils\u0026mdash;where the material is denser and retains more heat\u0026mdash;are shown in red, while the interior regions rapidly cool and appear blue or white. When using the fixed patching method described in Section \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e3.2\u003c/span\u003e, boundary-related temperature gradients and discontinuities often lead to prediction instability, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(b). For heavy-type wire rods, the lower temperature at the core of the image patch is still distinguishable from the background due to their thicker cross-sections and slower cooling rates. Conversely, fine-type wire rods cool rapidly, and within a single patch, the material temperature may vary significantly, resembling the background. Consequently, the segmentation model\u0026mdash;trained primarily on well-defined high-temperature material\u0026mdash;may underestimate or miss portions of the wire rod entirely when applied to fine-type images.\u003c/p\u003e\u003cp\u003eTo address this issue, we propose a sliding window-based patching method. Instead of dividing the image into four patches, we generated overlapping 128\u0026times;128 patches by shifting a window every 32 pixels in both horizontal and vertical directions. For each patch, an individual color filter was applied, normalizing the patch independently to capture local contrast better. Algorithm 1 outlines the classification process. Unlike the method described in Section \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e3.2\u003c/span\u003e, this approach allows up to 16 patches to contribute to the prediction of the same pixel (temperature) at a given location.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e\u003ccolgroup cols=\"1\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAlgorithm 1: Pixel Classification with Sliding Window Patches\u003c/p\u003e\u003cp\u003e# Input: Image I, Segmentation model M, Threshold θ\u003c/p\u003e\u003cp\u003e# Output: Final prediction map P\u003c/p\u003e\u003cp\u003efor each pixel (i, j) in I do\u003c/p\u003e\u003cp\u003epredictions \u0026larr; empty list\u003c/p\u003e\u003cp\u003efor each patch containing pixel (i, j) do\u003c/p\u003e\u003cp\u003epred \u0026larr; M.predict(patch)\u003c/p\u003e\u003cp\u003eappend pred[i, j] to predictions\u003c/p\u003e\u003cp\u003eend for\u003c/p\u003e\u003cp\u003eprediction_sum \u0026larr; sum of all values in predictions\u003c/p\u003e\u003cp\u003eif prediction_sum\u0026thinsp;\u0026ge;\u0026thinsp;θ then\u003c/p\u003e\u003cp\u003eP[i, j] \u0026larr; 1 // Classified as wire rod\u003c/p\u003e\u003cp\u003eelse\u003c/p\u003e\u003cp\u003eP[i, j] \u0026larr; 0 // Classified as background\u003c/p\u003e\u003cp\u003eend if\u003c/p\u003e\u003cp\u003eend for\u003c/p\u003e\u003cp\u003ereturn P\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe overlapping sliding window method is widely utilized in image processing tasks to effectively manage high-resolution images and ensure spatial continuity between adjacent patches (Chen et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). In this study, this approach addresses inconsistencies caused by rigid patch boundaries and enhances segmentation performance on fine-type wire rods. During the prediction phase, where the segmentation model trained in Section \u003cspan refid=\"Sec6\" class=\"InternalRef\"\u003e3.3\u003c/span\u003e is applied, a single pixel may be included in multiple overlapping patches, resulting in multiple predictions for that pixel. To resolve conflicts in such overlapping regions, threshold-based fusion techniques are commonly employed at either the window (patch) or pixel level (Fan et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Guan et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). In our implementation, we adopted a pixel-wise union-based threshold rule: if any prediction classifies a pixel as part of a wire rod, it is ultimately labeled as a wire rod. This deliberate overestimation strategy helps minimize the omission of true material, which is particularly beneficial in low-contrast regions. To aid comprehension, Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e illustrates the concept of sliding window-based semantic segmentation using 32 \u0026times; 32 patches (as opposed to the actual 128 \u0026times; 128), highlighting how overlapping regions contribute to a more stable semantic map.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(c), applying this method resolved classification gaps, particularly in the central double-ended patch that was previously misclassified in Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(b). Moreover, the semantic map exhibits an improved structural consistency, with clearly repeating circular patterns indicative of the wire rod coil structure. Nevertheless, despite overestimating the material areas, residual fragmentation remains, with dotted or broken contours still visible.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e4.2 Training the Restoration Model\u003c/h2\u003e\u003cp\u003eAlthough the enhanced segmentation map generated by the sliding window method presents a more complete structure, the results often appear visually fragmented to human observers familiar with the continuous curvature of wire rods. These dotted predictions resemble blue noise, introducing ambiguity in determining the true morphology of the wire rods.\u003c/p\u003e\u003cp\u003eTo enhance the prediction quality, a CNN model can be trained to infer the intact shape of a wire rod from imperfect semantic maps. We refer to this as the restoration model. Figure\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e provides an overview of the restoration process. Training data for this model were derived from heavy-type wire rod semantic maps, which provide more reliable structural patterns. Additionally, we gathered two types of input-output image patch pairs: semantic maps of manually annotated data from Section \u003cspan refid=\"Sec3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, and high-quality segments of the predicted semantic maps, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e. Using a sliding window with a stride of 8 pixels, we extracted 3,956 patches of 512 \u0026times; 512 pixels. Artificial noise was introduced by randomly setting 65% to 95% of the pixels in each input patch to zero, simulating the sparse prediction pattern seen in fine-type wire rod segmentation. The restoration model was trained to reconstruct the original shape from these degraded inputs. We used a standard 7:1:2 ratio to split the dataset into training, validation, and test sets.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e illustrates the network architecture used for restoration. The architecture of the restoration model is based on U-Net, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e, but differs from the semantic segmentation model (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e) in several key aspects, including input and output dimensions (512 \u0026times; 512\u0026times;1), the specific objective of noise suppression, and the morphological restoration of wire rod patterns.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e4.3 Results\u003c/h2\u003e\u003cp\u003eThe restoration model, trained to infer the complete morphology of the wire rods from incomplete semantic maps, achieved a test accuracy of 0.9632. This model is designed to reconstruct the full wire rod structure by learning from intentionally degraded input images and their corresponding intact versions, effectively addressing visual noise issues such as the \"blue noise\" effect observed in Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(c).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAs no labeled data exist for fine-type wire rods, quantitative evaluation is not feasible. Therefore, the semantic map must be assessed quantitatively through visual inspection. Figure\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(d) shows the results of applying both the sliding window method (Section \u003cspan refid=\"Sec10\" class=\"InternalRef\"\u003e4.2\u003c/span\u003e) and the restoration model (Section \u003cspan refid=\"Sec11\" class=\"InternalRef\"\u003e4.3\u003c/span\u003e) to fine-type wire rods. Compared with the original image in Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(a), which is visually ambiguous, the restored semantic map accurately predicts the expected wire rod regions. Notably, Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(d) demonstrates a significant improvement over Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(c). Although the latter contains noisy, fragmented predictions, the former exhibits smoother, continuous contours, resulting in a more realistic and coherent wire rod shape.\u003c/p\u003e\u003cp\u003eThis approach leverages the ability of the model to learn and reconstruct structural patterns, even from sparse or noisy inputs, thereby enhancing the overall visual quality and consistency of the wire rod segmentation. However, a limitation remains at the head region of the wire rod, where the prediction accuracy has not improved significantly. Despite this limitation, the current results effectively fulfill the primary objective highlighted in Chap.\u0026nbsp;4\u0026mdash;to facilitate a qualitative understanding of temperature dispersion across the entire wire rod length.\u003c/p\u003e\u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eThis study proposes a novel, automated method for acquiring post-cooling temperature data of wire rods using image-processing techniques, specifically semantic segmentation. The proposed methodology is designed to overcome challenges posed by the complex geometry of wire rods and limitations of conventional point-based temperature measurement methods in the steel industry.\u003c/p\u003e\u003cp\u003eOur approach consists of three components:\u003c/p\u003e\u003cp\u003e(1) a preprocessing technique that divides thermal images into manageable patches and enhances contrast using color filters;\u003c/p\u003e\u003cp\u003e(2) a U-Net-based semantic segmentation model trained to differentiate wire rods from the background; and\u003c/p\u003e\u003cp\u003e(3) a restoration model that reconstructs fine-type wire rod shapes from noisy predictions.\u003c/p\u003e\u003cp\u003eThe semantic segmentation model, trained on augmented heavy-wire rod data, achieved an accuracy of 0.9624, an F1-score of 0.9473, and an mIoU of 0.8998\u0026mdash;demonstrating its strong performance despite the intricate morphology of wire rods.\u003c/p\u003e\u003cp\u003eFor fine-type wire rods, we introduced a sliding window-based patch-generation method combined with a restoration model. This hybrid strategy mitigates prediction artifacts, connects fragmented segments, and generates more natural, continuous semantic maps, thereby indicating the applicability of the model to previously untrainable datasets.\u003c/p\u003e\u003cp\u003eThe significance of this study lies in its ability to automate full-length temperature profiling of wire rods, a task that was previously limited to manual spot-checking. This advancement has direct implications for quality control in wire rod production, with potential benefits including defect reduction and enhanced product uniformity.\u003c/p\u003e\u003cp\u003eFuture research should focus on acquiring annotated datasets for fine-type wire rods, enhancing the model through domain-specific data augmentation, and integrating the temperature profiling system with predictive models of physical properties, such as hardness prediction models (Pyo et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2024\u003c/span\u003e)\u0026mdash;to establish a comprehensive quality control framework for wire rod manufacturing.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003ea. Funding: This work was supported by the Korea Evaluation Institute of Industrial Technology (KEIT) and the Ministry of Trade, Industry, and Energy (MOTIE) of the Republic of Korea (grant number [RS-2022-00155473]: Development of technology for improving energy efficiency and product quality through the application of big data in the steel rolling process).\u003c/p\u003e\n\u003cp\u003eb. Competing Interests:\u0026nbsp;The authors report there are no competing interests to declare.\u003c/p\u003e\n\u003cp\u003ec. Data Availability Statement: Derived data supporting the findings of this study are available from the corresponding author on request.\u003c/p\u003e\n\u003cp\u003ed. Authors\u0026apos; contributions: All authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Seok-Kyu Pyo, Dong-Hee Lee, Sung-Jun Hur, Sang-Hyeon Lee, Sung-Jun Lim, Jong-Eun Lee. The first draft of the manuscript was written by Seok-Kyu Pyo and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAbraham, G. K. (2013). The Effect of an Encapsulated Specimen\u0026rsquo;s Mounting Material on Its Rockwell Hardness Test Results. \u003cem\u003eMetallography, Microstructure, and Analysis\u003c/em\u003e, \u003cem\u003e2\u003c/em\u003e(6), 378\u0026ndash;382. https://doi.org/10.1007/s13632-013-0104-6\u003c/li\u003e\n\u003cli\u003eAsthana, N., \u0026amp; Byeon, H. (2024). U-Net-based Recommender Systems for Political Election System using Collaborative Filtering Algorithms. \u003cem\u003eJournal of Information and Communication Convergence Engineering\u003c/em\u003e, \u003cem\u003e22\u003c/em\u003e(1), 7\u0026ndash;13. https://doi.org/10.56977/jicce.2024.22.1.7\u003c/li\u003e\n\u003cli\u003eChe, L., He, Z., Zheng, K., Xu, X., \u0026amp; Zhao, F. (2024). An automatic segmentation and quantification method for austenite and ferrite phases in duplex stainless steel based on deep learning. \u003cem\u003eJournal of Materials Chemistry A\u003c/em\u003e. https://doi.org/10.1039/d4ta05421k\u003c/li\u003e\n\u003cli\u003eChen, M., Zhao, X., \u0026amp; Xu, D. (2019). Image Inpainting for Digital Dunhuang Murals Using Partial Convolutions and Sliding Window Method. \u003cem\u003eJournal of Physics: Conference Series\u003c/em\u003e, \u003cem\u003e1302\u003c/em\u003e(3). https://doi.org/10.1088/1742-6596/1302/3/032040\u003c/li\u003e\n\u003cli\u003eDaigo, I., Murakami, K., Tajima, K., \u0026amp; Kawakami, R. (2023). Thickness Classifier on Steel in Heavy Melting Scrap by Deep-learning-based Image Analysis. \u003cem\u003eISIJ International\u003c/em\u003e, \u003cem\u003e63\u003c/em\u003e(1), 197\u0026ndash;203. https://doi.org/10.2355/isijinternational.ISIJINT-2022-331\u003c/li\u003e\n\u003cli\u003eFan, X., Riaz, I., Rehman, Y., \u0026amp; Shin, H. (2016). Vanishing point detection using random forest and patch-wise weighted soft voting. \u003cem\u003eIET Image Processing\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e(11), 900\u0026ndash;907. https://doi.org/10.1049/iet-ipr.2016.0068\u003c/li\u003e\n\u003cli\u003eGuan, M., Yu, W., Guo, Y., Huang, K., Zhang, J., Liang, K., \u0026amp; Ma, Z. (2024). Focus the Overlapping Problem on Few-Shot Object Detection via Multiple Predictions. \u003cem\u003eLecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)\u003c/em\u003e, \u003cem\u003e14426 LNCS\u003c/em\u003e, 427\u0026ndash;439. https://doi.org/10.1007/978-981-99-8432-9_34\u003c/li\u003e\n\u003cli\u003eHarste, K., \u0026amp; Wustner, E. (2004). Innovative solutions for the production of wire rod and bar. \u003cem\u003eSTAHL UND EISEN\u003c/em\u003e, \u003cem\u003e124\u003c/em\u003e(1), 43\u0026ndash;48.\u003c/li\u003e\n\u003cli\u003eHwang, J. K. (2020). Effect of Ring Configuration on the Deviation in Cooling Rate and Mechanical Properties of a Wire Rod during the Stelmor Cooling Process. \u003cem\u003eJournal of Materials Engineering and Performance\u003c/em\u003e, \u003cem\u003e29\u003c/em\u003e(3), 1732\u0026ndash;1740. https://doi.org/10.1007/s11665-020-04694-0\u003c/li\u003e\n\u003cli\u003eKoymatcik, H., Ahlatci, H., Sun, Y., \u0026amp; Turen, Y. (2018). Effect of carbon content and drawing strain on the fatigue behavior of tire cord filaments. \u003cem\u003eENGINEERING SCIENCE AND TECHNOLOGY-AN INTERNATIONAL JOURNAL-JESTECH\u003c/em\u003e, \u003cem\u003e21\u003c/em\u003e(3), 289\u0026ndash;296. https://doi.org/10.1016/j.jestch.2018.04.010\u003c/li\u003e\n\u003cli\u003eLaub, M., Bachmann I, B.-, Detemple, E., Scherff, F., Staudt, T., Mueller, M., Britz, D., Muecklich, F., \u0026amp; Motz, C. (2022). Determination of grain size distribution of prior austenite grains through a combination of a modified contrasting method and machine learning. \u003cem\u003ePRAKTISCHE METALLOGRAPHIE-PRACTICAL METALLOGRAPHY\u003c/em\u003e, \u003cem\u003e60\u003c/em\u003e(1), 4\u0026ndash;36. https://doi.org/10.1515/pm-2022-1025\u003c/li\u003e\n\u003cli\u003eLema, D. G., Pedrayes, O. D., Usamentiaga, R., \u0026amp; Garc\u0026iacute;a, D. F. (2023). Improved detection of subsurface defects through active thermography and ensembling techniques. \u003cem\u003eQuality Engineering\u003c/em\u003e, \u003cem\u003e0\u003c/em\u003e(0), 1\u0026ndash;17. https://doi.org/10.1080/08982112.2023.2177871\u003c/li\u003e\n\u003cli\u003eLong, J., Shelhamer, E., \u0026amp; Darrell, T. (2015). Fully convolutional networks for semantic segmentation. \u003cem\u003e2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) \u003c/em\u003e, 3431\u0026ndash;3440. https://doi.org/10.1109/CVPR.2015.7298965\u003c/li\u003e\n\u003cli\u003eMorales-Cervantes, A., Ch\u0026aacute;vez-Campos, G. M., Vergara-Hern\u0026aacute;ndez, H. J., Flores, J. J., \u0026amp; Guevara, E. (2023). Monitoring Steel Heating Processes Using Infrared Thermography and Deep Learning-Based Semantic Segmentation. \u003cem\u003eJOM\u003c/em\u003e. https://doi.org/10.1007/s11837-023-06212-8\u003c/li\u003e\n\u003cli\u003ePavlina, E. J., \u0026amp; Van Tyne, C. J. (2008). Correlation of Yield Strength and Tensile Strength with Hardness for Steels. \u003cem\u003eJOURNAL OF MATERIALS ENGINEERING AND PERFORMANCE\u003c/em\u003e, \u003cem\u003e17\u003c/em\u003e(6), 888\u0026ndash;893. https://doi.org/10.1007/s11665-008-9225-5\u003c/li\u003e\n\u003cli\u003ePedrayes, O. D., Lema, D. G., Usamentiaga, R., Venegas, P., \u0026amp; Garc\u0026iacute;a, D. F. (2022). Semantic segmentation for non-destructive testing with step-heating thermography for composite laminates. \u003cem\u003eMeasurement: Journal of the International Measurement Confederation\u003c/em\u003e, \u003cem\u003e200\u003c/em\u003e. https://doi.org/10.1016/j.measurement.2022.111653\u003c/li\u003e\n\u003cli\u003ePyo, S. K., Lee, D. H., Hur, S. J., Lee, S. H., Lim, S. J., Lee, J. E., \u0026amp; Moon, H. K. (2024). Prediction of spring steel wire rod hardness based on wire rod rolling process data: a case study. \u003cem\u003eInternational Journal of Advanced Manufacturing Technology\u003c/em\u003e, \u003cem\u003e133\u003c/em\u003e(1\u0026ndash;2), 1011\u0026ndash;1024. https://doi.org/10.1007/s00170-024-13740-3\u003c/li\u003e\n\u003cli\u003eRonneberger, O., Fischer, P., \u0026amp; Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation. \u003cem\u003eLecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)\u003c/em\u003e, \u003cem\u003e9351\u003c/em\u003e, 234\u0026ndash;241. https://doi.org/10.1007/978-3-319-24574-4_28\u003c/li\u003e\n\u003cli\u003eRukundo, O. (2023). Effects of Image Size on Deep Learning. \u003cem\u003eElectronics (Switzerland)\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(4). https://doi.org/10.3390/electronics12040985\u003c/li\u003e\n\u003cli\u003eTalebi, H., \u0026amp; Milanfar, P. (2021). Learning to Resize Images for Computer Vision Tasks. \u003cem\u003eCoRR\u003c/em\u003e, \u003cem\u003eabs/2103.09950\u003c/em\u003e. https://arxiv.org/abs/2103.09950\u003c/li\u003e\n\u003cli\u003eThomas, A., Durmaz, A. R., Straub, T., \u0026amp; Eberl, C. (2020). Automated quantitative analyses of fatigue-induced surface damage by deep learning. \u003cem\u003eMaterials\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(15). https://doi.org/10.3390/ma13153298\u003c/li\u003e\n\u003cli\u003eUsamentiaga, R., Garcia, D. F., \u0026amp; Perez, J. M. (2020). High-Speed Temperature Monitoring for Steel Strips Using Infrared Line Scanners. \u003cem\u003eIEEE Transactions on Industry Applications\u003c/em\u003e, \u003cem\u003e56\u003c/em\u003e(3), 3261\u0026ndash;3271. https://doi.org/10.1109/TIA.2020.2980478\u003c/li\u003e\n\u003cli\u003eWei, Y., Li, Y., Zhu, L., Liu, Y., Lei, X., Wang, G., Wu, Y., Mi, Z., Liu, J., Wang, H., \u0026amp; Gao, H. (2014). Evading the strength- ductility trade-off dilemma in steel through gradient hierarchical nanotwins. \u003cem\u003eNATURE COMMUNICATIONS\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e. https://doi.org/10.1038/ncomms4580\u003c/li\u003e\n\u003cli\u003eXie, L., Li, W., Fan, L., \u0026amp; Zhou, M. (2023). Automatic Identification of the Multiphase Microstructures of Steels based on ASPP-FCN. \u003cem\u003eSteel Research International\u003c/em\u003e. https://doi.org/10.1002/srin.202200204\u003c/li\u003e\n\u003cli\u003eXu, Z.-G., Wang, J., \u0026amp; Wang, L.-Y. (2018). Infrared Image Semantic Segmentation Based on Improved DeepLab and Residual Network. In \u003cem\u003e2018 10th International Conference on Modelling, Identification and Control (ICMIC)\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eYuan, X., Shi, J., \u0026amp; Gu, L. (2021). A review of deep learning methods for semantic segmentation of remote sensing imagery. In \u003cem\u003eExpert Systems with Applications\u003c/em\u003e (Vol. 169). Elsevier Ltd. https://doi.org/10.1016/j.eswa.2020.114417 \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"the-international-journal-of-advanced-manufacturing-technology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jamt","sideBox":"Learn more about [The International Journal of Advanced Manufacturing Technology](https://www.springer.com/journal/170)","snPcode":"170","submissionUrl":"https://submission.nature.com/new-submission/170/3","title":"The International Journal of Advanced Manufacturing Technology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"wire rod, thermal image, semantic segmentation, convolutional neural network","lastPublishedDoi":"10.21203/rs.3.rs-7625115/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7625115/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eHardness is a critical quality attribute of wire rods and is significantly influenced by cooling temperatures. To measure this temperature, infrared line scanners are typically employed to generate thermal images that provide full-length temperature data of wire rods. In these thermal images, some pixels correspond to the rod surface temperature, while others represent the background, which often contains noise. Accurately identifying rod-related pixels is essential, but this remains challenging due to the complex geometry of the wire rods. To address this, we propose a semantic segmentation method using a U-Net architecture to automatically distinguish wire rod pixels from the background. This method improves automation and eliminates the need for manual processing. We validated the approach on wire rods of varying diameters, from heavy-type (\u0026gt;\u0026thinsp;10 mm) to fine-type (\u0026le;\u0026thinsp;10 mm). For heavy-type rods, our method achieved an accuracy of 0.9624 and a mean Intersection over Union of 0.8998. For fine-type wire rods, where smaller temperature differences complicate segmentation, we introduced a restoration technique using a sliding window approach, which enhanced segmentation quality. This combined method enables automated full-length temperature measurement across different wire rod types, supporting improved quality control in wire rod manufacturing.\u003c/p\u003e","manuscriptTitle":"Semantic Segmentation of Steel Wire Rod Thermal Images for Automated Temperature Measurement","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-17 02:23:21","doi":"10.21203/rs.3.rs-7625115/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Major Revisions Needed","date":"2025-11-23T12:09:05+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"","date":"2025-10-06T16:54:37+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-10-05T19:12:21+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-09-18T06:12:20+00:00","index":"","fulltext":""},{"type":"submitted","content":"The International Journal of Advanced Manufacturing Technology","date":"2025-09-15T22:48:41+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"the-international-journal-of-advanced-manufacturing-technology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jamt","sideBox":"Learn more about [The International Journal of Advanced Manufacturing Technology](https://www.springer.com/journal/170)","snPcode":"170","submissionUrl":"https://submission.nature.com/new-submission/170/3","title":"The International Journal of Advanced Manufacturing Technology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"67306bfa-c768-4d86-bdf8-53f8259bfb54","owner":[],"postedDate":"October 17th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-03-16T16:00:50+00:00","versionOfRecord":{"articleIdentity":"rs-7625115","link":"https://doi.org/10.1007/s00170-026-17717-2","journal":{"identity":"the-international-journal-of-advanced-manufacturing-technology","isVorOnly":false,"title":"The International Journal of Advanced Manufacturing Technology"},"publishedOn":"2026-03-12 15:57:50","publishedOnDateReadable":"March 12th, 2026"},"versionCreatedAt":"2025-10-17 02:23:21","video":"","vorDoi":"10.1007/s00170-026-17717-2","vorDoiUrl":"https://doi.org/10.1007/s00170-026-17717-2","workflowStages":[]},"version":"v1","identity":"rs-7625115","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7625115","identity":"rs-7625115","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}