Mold Spot Detection for Paper Artifacts Based on Multimodal Feature Fusion

Mold Spot Detection for Paper Artifacts Based on Multimodal Feature Fusion | 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 Article Mold Spot Detection for Paper Artifacts Based on Multimodal Feature Fusion Xuexu Deng, Ya Zhao, Dan Qin, Zheng Ma, Zhongyu Xiao, Xiling Luo, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6292446/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 25 Oct, 2025 Read the published version in npj Heritage Science → Version 1 posted 11 You are reading this latest preprint version Abstract As an important carrier of Chinese civilization, the prevention and control of mold growth in paper cultural relics is the core challenge in the field of cultural relic protection. To address the issues of low sensitivity and poor timeliness of traditional detection methods, this study proposes a TriplePath Multimodal Feature Fusion Network (TPMFN) based on hyperspectral imaging. Our research constructs an improved two-dimensional convolutional network (2D-CNN), extracts RGB bands using spatial attention mechanism, and introduces improved SpectralFormer module to accurately capture mold specific responses in the 400-1000nm spectral range, ultimately achieving efficient integration and classification decision-making of multi-source information. In this study six categories of typical fungal colonization samples from paper-based cultural relics were analyzed. The research results indicate that the proposed TPMFN has significant advantages over baseline models including including Support Vector Machine (SVM), 1D-CNN, 2D-CNN, SSFTT, HybridSN and SpectralFormer, achieving 98.84% overall accuracy and 98.54% kappa coefficient. Notably, it attained a 3.6% improvement in detailed feature identification accuracy compared to the suboptimal baseline. The ablation study with component-wise evaluation revealed that the dual-attention mechanism enhanced feature discriminative power with a relative increase of 7.5%. This study provides a high-precision and practical solution for mold detection of paper cultural relics, which has important practical value for establishing a preventive protection system Hyperspectral imaging Mold spot detection Multimodality Feature fusion Deep learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 1. Introduction As pivotal witnesses to the splendor of Chinese civilization, paper-based cultural heritage artifacts have profoundly shaped intercultural dialogue and promoted the evolution of global civilizations. From classical paintings and canonical texts to epistolary documents and archival records, these artifacts offer invaluable research resources through their substantive content and refined sheet-forming techniques [ 1 ] . Nevertheless, the inherent hygroscopic fragility of paper matrices renders them susceptible to biodegradative and environmental stressors during storage and display processes, thereby escalating the complexity of preventive conservation interventions. Paper is one of the most vulnerable materials to mold growth due to its high content of cellulose, hemicellulose, and other nutrients. Mold secretes cellulase to degrade and absorb these components. During this process, the structure of the paper gradually deteriorates, and the bonds between fibers weaken, leading to a decline in its mechanical strength [ 2 ] . Furthermore, this biodegradation process may be accompanied by a series of complex chemical reactions, which further accelerate the aging and embrittlement of the paper, thereby severely damaging the historical and artistic value of cultural relics [ 3 – 5 ] . Current mold remediation approaches for cultural relics are primarily categorized into three modalities: mechanical, physical, and biochemical methods [ 6 ] . Mechanical interventions employ tools such as soft brushes, conservation scalpels, and HEPA-filtered vacuum systems to remove microbial colonization, though complete microbial elimination remains unattainable [ 7 ] . Physical techniques including UV irradiation and laser ablation demonstrate efficacy in microbial decontamination, yet show limited effectiveness against most bacterial strains and dematiaceous fungi [ 8 ] . While biochemical approaches prove more potent, the taxonomic diversity of molds necessitates species-level identification to devise targeted conservation strategies that prevent secondary damage, as monotherapeutic interventions often fail to achieve comprehensive eradication [ 9 ] .Therefore, the precise identification of fungal species serves not only as a prerequisite for mold remediation, but more crucially as a pivotal measure in safeguarding paper heritage artifacts against prolonged fungal colonization. Current conventional detection methodologies for paper-based cultural relics encompass morphological characterization, molecular biology techniques [ 10 ] , and biochemical assays [ 11 ] . However, these approaches rely solely on singular technological modalities - be they image-based or spectroscopic - which impedes comprehensive mycological identification and consequently manifests inherent limitations. Hyperspectral imaging technology, as an advanced detection means fusing image and spectral information, has shown a wide range of application prospects in the field of mold detection with its high sensitivity, multi-dimensional information acquisition capability and non-destructive detection characteristics. At present, this technology is mainly applied in the fields of food quality assessment [ 12 – 13 ] , crop disease detection [ 14 – 15 ] and preliminary progress has been made for the identification and assessment of mold contamination in the field of cultural relics protection.For example, Lu [ 16 ] et al. proposed an automated labeling method based on hyperspectral images for the detection of mold damage in murals, which effectively addressed the issues of time consumption, subjectivity, and inconsistency in traditional manual labeling methods. By integrating the spatial and spectral information of hyperspectral images, accurate identification and labeling of mold on mural surfaces were achieved, demonstrating high practical value and research significance. Dai [ 17 ] et al. systematically studied the spectral characteristics of simulated samples of paper artifacts affected by foxing based on hyperspectral imaging technology. By applying band arithmetic and the minimum noise fraction (MNF) method, the ability to extract and differentiate features of the infected areas was effectively enhanced, and a discriminative model based on the K-nearest neighbor method and BP neural network was constructed, achieving an outstanding accuracy of over 79% in foxing detection.The above research provides important technical support for the accurate detection and scientific control of mold contamination and further extends the application potential of hyperspectral imaging technology in cultural heritage conservation. In this study, we integrate digital image data with hyperspectral imaging data, employing multimodal data fusion technology and machine learning methodologies to achieve efficient detection of microbial colonization on paper-based cultural heritage. Through evaluating the impact of different feature extraction methods on mold identification and comparing their accuracy and robustness, we establish an optimal detection model. Furthermore, we extend the applicability of multimodal data fusion technology in cultural heritage conservation. This approach not only provides an effective solution for early-stage detection of biodeterioration in paper artifacts, but also pioneers new methodologies for multimodal data fusion in heritage science, demonstrating substantial scientific and practical significance. 2. Experiment section 2.1 Simulated mold infestation samples To ensure the diversity and accuracy of the experimental data, this study selected cotton-linen raw Xuan paper, cut into 5×5 cm samples, as the base material for simulating mold contamination. Based on literature from the past five years on common mold species found in the preservation environments of paper-based cultural relics, six representative mold species from different genera were chosen as experimental subjects, including Aspergillus niger , Penicillium citrus , Trichoderma longibrachiatum , Alternaria alternata , Paecilomyces lilacinus and Cladosporium cladosporioides . These six mold species belong to distinct genera, reflecting both the biodiversity and ecological adaptability of molds. Widely distributed in natural environments, they can potentially cause deterioration in cultural relics. Therefore, it is essential to prioritize monitoring these genera during mold detection and conservation processes. The paper substrates and fungal strains used in this experiment were supplied by the Key Scientific Research Base for Cultural Relic Pest Control under the State Administration of Cultural Heritage. A photograph of the artificially inoculated mold samples is presented in Fig. 1. 2.2 Mold stain acquisition system for paper-based cultural relics In this study, a hyperspectral image acquisition system for mold stains on paper-based cultural relics was built, as shown in Fig. 2 . The system consists of an iSpecHyper-VS1000 portable hyperspectral imager from Lyson Optical, two halogen light sources with adjustable brightness and color temperature, and a Canon RF 24mm focal length lens. The specific parameters of the hyperspectral imager are listed in Table 1 . After multiple imaging tests, the optimal parameters were determined as follows: a focal length of 4615mm, a frame rate of 16, and an integration time of 56338. Images captured with these parameters exhibited excellent quality, with sharp edges and minimal distortion. Table 1 iSpecHyper-VS1000 hyperspectral imager parameters Main Technical Specifications Specification Parameters Spectral Range 400 ~ 1000 Spectral Resolution <3 F-number F/2.6 Imaging method Pushbroom Imaging Number of spectral channels 300 Detector COMS/InGaAs(TE Cooled) AD Dynamic Range 12 Frame Rate of Spectral Camera 50 Spatial Resolution of Image (Pixels) 1920×1200 Detector Pixel Size 5.86*5.86 2.3 Experimental setup 1) Evaluation metrics In order to verify the validity of the models, this paper quantitatively analyzes the classification results of each model using three commonly used classification performance assessment metrics: Overall Accuracy ( OA ), Average Accuracy ( AA ), and Kappa Coefficient ( Kappa ). Among them, OA reflects the overall classification accuracy of the model across all categories, AA indicates the mean classification accuracy for each category, and the Kappa coefficient measures the consistency between the classification results and random classification results, thus providing a more comprehensive evaluation of the model's performance. Meanwhile, to further visualize and compare the classification performance of each model, this paper also conducts qualitative analysis by visualizing classification maps to observe the spatial distribution characteristics and the accuracy of the classification boundaries, in order to comprehensively evaluate the classification performance of the models from both quantitative and qualitative dimensions. 2) Comparative analysis of models In this paper, several state-of-the-art models are selected for comparative analysis with the proposed model, including: SVM, 1-D-CNN [ 18 ] , 2-D-CNN [ 19 ] , SSFTT [ 20 ] , SpectralFormer [ 21 ] , and HybridSN [ 22 ] . To ensure the uniformity and validity of the experimental data, a unified preprocessing approach is applied, where the Savitzky-Golay(SG) smoothing filter is used to remove noise, and principal component analysis(PCA) is employed to reduce the number of spectral bands to 30 dimensions. This helps eliminate redundant information and highly similar features, thereby improving recognition accuracy. 3) Experimental setup In order to validate the multimodal feature fusion-based classification algorithm for hyperspectral mold detection proposed in this paper, all experiments were conducted using Python 3.12 and implemented in the PyCharm 2024.3.1 integrated development environment. The experiments were performed on a high-performance personal computer equipped with an Intel(R) Core i5-14600KF processor, an NVIDIA GeForce RTX 4070 GPU, and 32 GB of RAM. During the experiments, all deep learning models employed Cross-Entropy Loss as the optimization objective, in conjunction with the Adam optimizer for parameter updates. Stochastic Gradient Descent (SGD) was used for model training optimization. The learning rate was uniformly set to 0.001, and the number of training epochs was fixed at 50 to ensure model stability and convergence, thereby enabling a reliable evaluation of each model’s performance. 3. Related Work In this section, we first introduce the TPMFN model proposed in this paper, followed by a detailed explanation of the design logic for feature extraction along the three different paths within the model and their roles in multimodal feature extraction and fusion. Finally, we summarize the advantages and potential of the TPMFN model. Here, after the data block division, the input hyperspectral image features are denoted as , where B represents the batch size, H and W represent the spatial dimensions, and C represents the number of channels. 3.1 TPMFN The model aims to enhance the accuracy and robustness of mold stain detection through the efficient fusion of multimodal features. The core design concept of the TPMFN model is to fully exploit the characteristics of three different modalities for feature extraction: the spectral dimension of hyperspectral data, the joint spatial-spectral dimension, and the spatial dimension of RGB images. By employing a deep fusion mechanism, the model achieves a comprehensive representation of global features. The network structure diagram is shown in Fig. 3 . In the TPMFN model, the input hyperspectral image is first processed to extract three RGB bands, generating an RGB digital image. Then, different feature extraction methods are applied to different modalities.For the RGB digital modality, 2D convolution neural network(CNN) is employed to capture spatial characteristics such as edge morphology and color variations in the pseudo-color image. Subsequently, a spatial attention module is introduced to enhance the spatial texture representation of the mold stain region while filtering out interference from non-mold stain areas.For the hyperspectral data modality, feature extraction follows two paths. In spectral feature extraction, a Spectral Transformer is first used to capture subtle spectral differences of mold stains across hyperspectral bands. Then, Spectral Attention is introduced to emphasize key spectral bands, generating spectral features.For spatial-spectral feature extraction, a Hybrid Convolutional Network is employed to jointly analyze the spatial distribution and spectral variations of mold stains in hyperspectral data. 1) Spatial features The 2D CNN is a deep learning model specialized in processing 2D data. Its core mechanism utilizes convolutional operations for local feature extraction, where parameter-shared kernels capture low-level features (e.g., edges, textures, structures) within localized spatial regions, while progressively learning higher-level semantic patterns through layered hierarchies. In this work, we first extract and flatten the three RGB bands from input X. A three-stage 2D convolutional network then hierarchically extracts spatial features from the hyperspectral image, with optimization applied to features at each layer to amplify discriminative spatial regions. The flattened outputs are subsequently fed into a three-layer fully connected network for progressive dimensionality reduction, ultimately generating category predictions. The spatial attention mechanism enhances critical regions by computing spatial weight distributions, with the attention energy defined in Eq. ( 1 ): In Eq. ( 1 ), Q is the query matrix, and K is the key matrix. Then, a Softmax operation is applied to each row of E to obtain the attention weight A. The formula for calculating the attention weight is defined in Eq. (2) In Eq. (2), A ij represents the weight distribution of the pixel spatial position at the i-th row and j-th column. The specific network structure diagram of the spatial feature extraction algorithm is shown in Fig. 4. Figure 4. The network structure diagram of the spatial feature extraction model 2) Spectral features SpectralFormer is an improved Transformer-based model that enables local spectral representations from multiple neighboring bands at each encoding location. This is achieved through the Group Smart Spectral Embedding (GSE), which enhances the capture of fine spectral variations, and the Cross-Layer Adaptive Fusion (CAF), which improves the transfer of information between layers. In addition, SpectralFormer incorporates a cross-layer skip connection, which adaptively learns to fuse their "soft" residuals, gradually passing memory-like components from shallow to deep layers.Although SpectralFormer is good at capturing global sequence information, it is less effective at modeling the local contextual information of hyperspectral data. To address this limitation, this paper proposes an improved spectral feature extraction model based on SpectralFormer. Its structure is shown in Fig. 5 . First, the spectral attention mechanism is introduced, performing global average pooling to obtain the global representation of spectral channels. It then assigns weights to spectral bands to emphasize important ones, followed by applying linear projection for spectral embedding. Classification tokens and positional embeddings are then added, followed by sequence modeling using the Transformer module, and finally, the classification results are output.The input is the extracted spectral information where N represents the number of features. The spectral attention process is formulated as shown in Eq. ( 3 ): 3) Spatial-spectral features The hybrid convolutional network employs multi-dimensional feature extraction by applying 3D convolution to concurrently extract spectral and spatial features from hyperspectral data, while utilizing 2D convolution to further extract spatial features. The network then progressively fuses features across layers, integrating high-dimensional spectral features into low-dimensional representations and introducing 2D convolution operations to aggregate and optimize spatial feature representations. This architecture demonstrates significant advantages in effectively integrating multi-dimensional features from hyperspectral data, thereby enhancing both the accuracy and robustness of classification tasks. By adopting a hierarchical feature extraction strategy, it optimizes computational resource allocation to address the inherent complexity and redundancy challenges in hyperspectral data processing. The detailed structure of the hybrid convolutional network model employed in this study is illustrated in Fig. 6. 3.2 TPMFN model advantages To address the limitations of the aforementioned models, the TPMFN proposed in this paper exhibits the following advantages: 1. Multi-path architecture with adaptive receptive fields: By integrating convolution operations with varying receptive field sizes (kernel sizes), the TPMFN effectively captures multi-scale spatial-spectral features in hyperspectral images, simultaneously enhancing representation capabilities for both fine-grained details and global contextual patterns. 2. Dual-branch complementary learning: The architecture comprises two specialized branches dedicated to spectral feature extraction and spatial feature analysis, respectively. This design enables concurrent focus on localized spectral variations and macroscopic spatial distribution characteristics in hyperspectral imagery. 3. Hierarchical feature fusion: Through a learnable fusion strategy, the TPMFN achieves feature map alignment and cross-modality integration after multi-path extraction. This mechanism produces comprehensive characterizations of hyperspectral data complexity, ultimately leading to superior classification performance. 4. Results and Discussion In this section, we first describe the hyperspectral image dataset used in our experiments. Subsequently, ablation experiments are conducted to systematically analyze the impacts of three distinct spectral-spatial feature extraction modules on model performance. Finally, the proposed model is benchmarked against mainstream approaches, with both quantitative metrics and visual comparisons being utilized to evaluate fungal lesion classification capabilities across different models. 4.1 Data-set description The hyperspectral image dataset used in this study contains mold spots induced by the six aforementioned fungal infections, acquired at a sampling height of 40 cm. Each image has dimensions of 521 × 364 pixels, containing 19,951 valid pixels per spectral band across 300 spectral channels. The dataset was partitioned into training(10%) and test(90%) subsets. Table 2 details the class nomenclature and corresponding sample distribution between the training and test sets for the classification task. Table 2 Class-specific mold coverage characteristics and training-test sample allocation No. Class Training Test 1 Paecilomyces lilacinus 250 2251 2 Aspergillus niger 358 3225 3 Penicillium citrus 365 3289 4 Alternaria alternata 652 5867 5 Trichoderma longiformis 184 1650 6 Cladosporium cladosporioides 186 1674 Total 1995 17956 4.2 Ablation study In order to fully validate the effectiveness of the proposed method, ablation experiments with different component combinations were conducted on the dataset mentioned in this paper. Six configurations were considered, and the impact of each component on the overall model accuracy was analyzed through classification performance. All experimental results are listed in Table 3 . Specifically, the model was divided into four modules: 2D-CNN + Spectral_Att, SpectralFormer, Spectral_Att, and Hybrid Network. Case 1 (excluding both SpectralFormer and Spectral_Att) achieved the lowest accuracy of 90.31%. Case 4 (without 2D-CNN + Spectral_Att) showed a slight improvement to 91.01%. Case 5 (removing Hybrid Network) reached 93.45% accuracy by relying solely on independent processing of spatial and spectral features. Comparing Case 3 (SpectralFormer alone) and Case 6 (SpectralFormer + Spectral_Att), the accuracy significantly increased to 97.12%. This demonstrates that Spectral_Att positively enhances spectral feature processing, thereby improving classification accuracy. Table 3 Presents the analysis of the proposed model conducted on this dataset (Suboptimal results) No. Component Indicators 2D-CNN Spatial_Att SpectralFormer Spectral_Att Hybrid Network OA(%) AA(%) ×100 1 √ × × √ 90.31 85.25 87.68 2 √ × √ √ 93.58 91.04 92.83 3 √ √ × √ 92.91 90.19 92.27 4 × √ √ √ 91.01 88.81 89.58 5 √ √ √ × 93.45 90.39 91.46 6 √ √ √ √ 97.12 93.24 95.16 To further validate the effectiveness of the proposed algorithm, thorough comparative experiments employing different feature fusion approaches were conducted on this data-set. Three distinct fusion methods - additive fusion, multiplicative fusion, and concatenation fusion - were implemented, with all experimental results summarized in Table 4 . As shown in Fig. 7 , it can be observed that models employing multiplicative fusion generally demonstrate superior classification performance compared to those using additive fusion and concatenation fusion, achieving a top accuracy of 98.84%.This indicates that the multiplicative fusion approach exhibits enhanced capability in feature interaction and information coupling, enabling more effective exploitation of complementary relationships among multimodal features to improve classification performance. Table 4 Performance analysis of the proposed model on the benchmark dataset (Bold entries denote optimal results) No. Fusion Methods OA(%) AA(%) ×100 1 Additive Fusion 94.69 91.65 93.26 2 Multiplicative Fusion 98.84 98.52 98.54 3 Concatenation Fusion 97.46 96.67 96.81 4.3 Quantitative analysis Table 5 presents the OA, AA, Kappa coefficient, and per-class classification accuracies obtained by all methods described in Section 3 . The optimal results are highlighted in bold.The evaluation data clearly demonstrate that the proposed TPMFN method achieves the best performance, yielding the highest OA, AA, and Kappa coefficient values, along with superior classification accuracies for specific categories. For instance, in the Trichoderma longibrachiatum class, models including SVM, 1D-CNN, 2D-CNN, Spectralformer, and SSFTT exhibit limited effectiveness, potentially due to the small sample size and non-concentrated distribution of this class, which hinders feature learning. Furthermore, the percentage-based random sampling strategy may exacerbate class imbalance issues. In contrast, TPMFN delivers consistently high classification accuracies (above 96%) across all classes, indicating its strong capability in handling class imbalance for hyperspectral image classification tasks. Whether for dominant or minor classes, the model maintains robust performance. This consistency likely stems from its multimodal feature fusion mechanism and efficient hierarchical feature extraction. However, there is another case: in the category of Paecilomyces lilacinus , the classification accuracy of SSFTT is significantly higher than that of TPMFN. The main reason is the highly convergent sample distribution of this category, which forms a compact, nearly circular pattern, whereas the sample distributions of other categories are more dispersed.Therefore, the proposed method does not show significant advantages in classifying this category. However, for datasets with discrete and localized sample points, TPMFN demonstrates a significant advantage, enabling better capture of fine-grained local information.This paper also investigates the impact of different training sample proportions on different models, as shown in Fig. 8 . With a small number of samples, TPMFN still maintains good performance. As the sample size increases, the performance of SSFTT and HybridSN is only slightly lower than that of the proposed method. Table 5 Classification accuracy of different classification methods on the dataset (Bold data indicates the best results in each category) 序号 SVM 1-D-CNN 2-D-CNN SSFTT Spectralformer HybirdSN TPMFN 1 34.18 46.38 85.66 98.58 88.31 83.65 97.46 2 44.65 59.50 87.13 98.75 85.74 97.51 99.06 3 60.18 83.37 92.54 97.63 99.05 96.89 100.00 4 74.96 92.39 93.16 97.92 98.46 98.95 99.45 5 27.79 34.01 61.40 58.64 63.52 85.33 96.54 6 25.17 62.43 48.85 96.10 97.05 99.28 99.16 OA(%) 61.18 70.90 85.09 94.39 91.71 95.38 98.84 AA(%) 54.47 61.07 78.45 90.35 88.78 93.71 98.52 ×100 56.09 62.87 80.94 92.90 89.53 94.17 98.54 4.4 Visual evaluation The classification maps of the aforementioned methods on this dataset are presented in Fig. 9. As observed, the TPMFN classification map exhibits the cleanest spatial patterns and closest resemblance to the ground truth. In contrast, conventional methods including SVM, 1D-CNN and 2D-CNN demonstrate limited capability in capturing discriminative spectral signatures or deep spatial features, resulting in noisy outputs with substantial misclassifications. This fundamental limitation ultimately leads to their poor performance in target recognition tasks.While advanced approaches like SpectralFormer, SSFTT and HybridSN achieve classification accuracies exceeding 90%, persistent errors remain in complex scenarios, particularly within regions containing Paecilomyces lilacinus and Trichoderma longibrachiatum species. Remarkably, the proposed TPMFN method successfully identifies these challenging mold colonies with significantly higher accuracy, quantitatively validating its superior performance. 5. Conclusion In this paper, we propose a multimodal feature fusion method, TPMFN, based on hyperspectral imaging to address the accuracy and efficiency challenges in mold detection in paper artifacts. TPMFN integrates spectral, spatial, and joint spatial-spectral features within a multipath structure to achieve deep feature extraction and fusion.Specifically, TPMFN captures spectral details using a Spectral Transformer, extracts spatial patterns of mold spots via a combination of 2-D CNN and a spatial attention mechanism, and further integrates multi-dimensional features through a hybrid convolutional network to enhance feature representation capability and improve computational efficiency.To further explore this capability, we investigate several fusion modules, including additive, multiplicative, and concatenation-based fusion. Experimental results demonstrate that the multiplicative fusion method significantly outperforms traditional classification methods and single-path networks in terms of accurately capturing local feature variations of mold spots, enhancing key information regions, and improving classification performance.The innovative design of TPMFN offers a novel perspective for hyperspectral image processing tasks and provides an efficient and intelligent solution for mold detection in paper artifacts. Declarations Funding This work was financially supported by the Chongqing Talents Program [grant numbers cstc2021ycjh-bgzxm0287]. Conflicts of Interest The authors declare no conflict of interest. Author Contribution Deng Xuexu, Zhao Ya and Qin Dan contributed to the study conception and design.Material preparation , data collection and analysis were preformed by Ma zheng, Xiao Zhongyu, Luo Xiling, Song Tao, Tang Bin and Wang Jianxu. Zhao Mingfu and Tang Huan are responsible for the supervision of research activities.The first draft of the manuscript was written by Deng Xuexu.All authors read and approved the final manuscript. Acknowledgement Special thanks to the key scientific research base of the State Administration of Cultural Heritage for research on pest control of cultural relics in collections (China Three Gorges Museum, Chongqing) for providing samples for this paper. References Zhang, Xu, et al. "Chemistry directs the conservation of paper cultural relics." Polymer Degradation and Stability 207 (2023): 110228. Carter, Henry A. "The chemistry of paper preservation: part 2. The yellowing of paper and conservation bleaching." Journal of Chemical Education 73.11 (1996): 1068. Pinheiro, Ana Catarina, Sílvia Oliveira Sequeira, and Maria Filomena Macedo. "Fungi in archives, libraries, and museums: a review on paper conservation and human health." Critical reviews in microbiology 45.5-6 (2019): 686-700. Yang Liguang. Causes of Damage to Paper Artifacts and Measures for Their Protection and Restoration [J]. Collection, 2023, (12): 110-112(in Chinese). Schmitz, Kevin, et al. "Preserving cultural heritage: Analyzing the antifungal potential of ionic liquids tested in paper restoration." PloS one 14.9 (2019): e0219650. Wu Fasi, Li Jie, Li Ruixi, et al. Research Progress on the Application of Fungicides and Antibacterial Nanomaterials in Cultural Heritage Conservation [J]. Sciences of Conservation and Archaeology, 2023, 35(05): 115-127(in Chinese). Rivas T, Pozo-Antonio J S, de Silanes M E L, et al. Laser versus scalpel cleaning of crustose lichens on granite[J]. Applied Surface Science, 2018, 440: 467-476. Pfendler S, Einhorn O, Karimi B, et al. UV-C as an efficient means to combat biofilm formation in show caves: evidence from the La Glacière Cave (France) and laboratory experiments[J]. Environmental Science and Pollution Research, 2017, 24: 24611-24623. Meng, Qingxia, et al. "A biological cleaning agent for removing mold stains from paper artifacts." Heritage Science 11.1 (2023). Yan Li, Hong Wei, Lü Xiaofang. Isolation and Molecular Identification of Molds on Archived Paintings and Silk Artifacts [J]. Cultural Relics Identification and Appreciation, 2021, (10): 60-65. Strycker B D, Han Z, Duan Z, et al. Identification of toxic mold species through Raman spectroscopy of fungal conidia[J]. PloS one, 2020, 15(11): e0242361. Vashpanov Y, Heo G, Kim Y, et al. Detecting green mold pathogens on lemons using hyperspectral images[J]. Applied Sciences, 2020, 10(4): 1209. Chun S W, Song D J, Lee K H, et al. Deep learning algorithm development for early detection of Botrytis cinerea infected strawberry fruit using hyperspectral fluorescence imaging[J]. Postharvest Biology and Technology, 2024, 214: 112918. Zou Z, Zhen J, Wang Q, et al. Research on nondestructive detection of sweet-waxy corn seed varieties and mildew based on stacked ensemble learning and hyperspectral feature fusion technology[J]. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 2024, 322: 124816. Siripatrawan U, Makino Y. Assessment of food safety risk using machine learning-assisted hyperspectral imaging: Classification of fungal contamination levels in rice grain[J]. Microbial Risk Analysis, 2024, 27: 100295. Lu M, Zhang Q, Li Y, et al. Research on Labeling Method for Fungal Diseases in Murals Based on Hyperspectral Images[C]//2024 3rd International Conference on Image Processing and Media Computing (ICIPMC). IEEE, 2024: 93-100. Dai R, Tang B, Zhao M, et al. Study on Detection Method of Foxing on Paper Artifacts Based on Hyperspectral Imaging Technology[C]//Journal of Physics: Conference Series. IOP Publishing, 2021, 2010(1): 012177. Hu W, Huang Y, Wei L, et al. Deep convolutional neural networks for hyperspectral image classification[J]. Journal of Sensors, 2015, 2015(1): 258619. Zhao W, Du S. Spectral–spatial feature extraction for hyperspectral image classification: A dimension reduction and deep learning approach[J]. IEEE Transactions on Geoscience and Remote Sensing, 2016, 54(8): 4544-4554. Sun L, Zhao G, Zheng Y, et al. Spectral–spatial feature tokenization transformer for hyperspectral image classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-14. Hong D, Han Z, Yao J, et al. SpectralFormer: Rethinking hyperspectral image classification with transformers[J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 60: 1-15. Roy S K, Krishna G, Dubey S R, et al. HybridSN: Exploring 3-D–2-D CNN feature hierarchy for hyperspectral image classification[J]. IEEE Geoscience and Remote Sensing Letters, 2019, 17(2): 277-281. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 25 Oct, 2025 Read the published version in npj Heritage Science → Version 1 posted Editorial decision: Revision requested 13 May, 2025 Reviews received at journal 13 May, 2025 Reviews received at journal 12 May, 2025 Reviewers agreed at journal 23 Apr, 2025 Reviewers agreed at journal 23 Apr, 2025 Reviews received at journal 23 Apr, 2025 Reviewers agreed at journal 07 Apr, 2025 Reviewers invited by journal 30 Mar, 2025 Editor assigned by journal 26 Mar, 2025 Submission checks completed at journal 26 Mar, 2025 First submitted to journal 24 Mar, 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-6292446","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":444207273,"identity":"3a31d1c7-29d0-429d-8031-6d967e1614cc","order_by":0,"name":"Xuexu Deng","email":"","orcid":"","institution":"Chongqing University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Xuexu","middleName":"","lastName":"Deng","suffix":""},{"id":444207276,"identity":"3a27f7c6-af3e-4bdf-ab26-6178d4a1ee2b","order_by":1,"name":"Ya Zhao","email":"","orcid":"","institution":"Chongqing Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Ya","middleName":"","lastName":"Zhao","suffix":""},{"id":444207281,"identity":"f1250d83-f473-4491-a158-216589a0652e","order_by":2,"name":"Dan Qin","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA3UlEQVRIiWNgGAWjYDACZgYGCTjng4GNHQla2BgYGGcUpCUTZRFcCzPPh0OMDYSU8x3nPXjj5w4Gef757Vc32xgcYGZgP3x0Az4tkof5ki17zzAYzjjGU3Y7x+AOHwNPWtoNfFoMDvOYSfC2MSQYsPGkAbU8Y2aQ4DEjqEXyL0yLhcFhxgZitEhDbGE/dpuBGC2Sh3mMrWXbQH7JYbvZY5CWzEbIL3znzxjefNsGDLHm489u/PhjY8fPfvgYXi0MB8DkfyDmMQAz2fAqR2gBAfYHBBWPglEwCkbByAQAZghGSrzr5VoAAAAASUVORK5CYII=","orcid":"","institution":"Research on Harmful Biological Control of Cultural Relics in Collections Key Scientific Research Base of the State Administration of Cultural Heritage (Chongqing China Three Gorges Museum)","correspondingAuthor":true,"prefix":"","firstName":"Dan","middleName":"","lastName":"Qin","suffix":""},{"id":444207282,"identity":"3160a332-99dc-4611-bc60-42a2fd7ea435","order_by":3,"name":"Zheng Ma","email":"","orcid":"","institution":"Research on Harmful Biological Control of Cultural Relics in Collections Key Scientific Research Base of the State Administration of Cultural Heritage (Chongqing China Three Gorges Museum)","correspondingAuthor":false,"prefix":"","firstName":"Zheng","middleName":"","lastName":"Ma","suffix":""},{"id":444207283,"identity":"23998ff4-0ec4-451d-9d3a-30e14d917026","order_by":4,"name":"Zhongyu Xiao","email":"","orcid":"","institution":"Chongqing University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Zhongyu","middleName":"","lastName":"Xiao","suffix":""},{"id":444207284,"identity":"2abdd1c1-2a45-4eb7-9f94-0d79f9e36605","order_by":5,"name":"Xiling Luo","email":"","orcid":"","institution":"Chongqing University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Xiling","middleName":"","lastName":"Luo","suffix":""},{"id":444207285,"identity":"67bd75e1-1b47-4b3e-a7a8-b6cd97a50287","order_by":6,"name":"Mingfu zhao","email":"","orcid":"","institution":"Chongqing University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Mingfu","middleName":"","lastName":"zhao","suffix":""},{"id":444207286,"identity":"77a13f58-306f-4b30-963b-09bcf0ea5c92","order_by":7,"name":"Tao Song","email":"","orcid":"","institution":"Chongqing University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Tao","middleName":"","lastName":"Song","suffix":""},{"id":444207287,"identity":"ae68aab9-2005-4294-81cc-73129e9ef374","order_by":8,"name":"Jianxu Wang","email":"","orcid":"","institution":"Chongqing University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Jianxu","middleName":"","lastName":"Wang","suffix":""},{"id":444207288,"identity":"22b6448a-fe77-4a0e-abc2-7b6656323f9c","order_by":9,"name":"Bin Tang","email":"","orcid":"","institution":"Chongqing University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Bin","middleName":"","lastName":"Tang","suffix":""},{"id":444207289,"identity":"e335d750-be90-40da-88a1-75855b37a403","order_by":10,"name":"Huan Tang","email":"","orcid":"","institution":"Research on Harmful Biological Control of Cultural Relics in Collections Key Scientific Research Base of the State Administration of Cultural Heritage (Chongqing China Three Gorges Museum)","correspondingAuthor":false,"prefix":"","firstName":"Huan","middleName":"","lastName":"Tang","suffix":""}],"badges":[],"createdAt":"2025-03-24 07:08:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6292446/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6292446/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s40494-025-02102-1","type":"published","date":"2025-10-25T16:17:05+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":81024333,"identity":"ae5cf3c9-9c15-45ff-a022-cb5ac67d9960","added_by":"auto","created_at":"2025-04-21 10:15:21","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1122142,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eImages of simulated mold spots: (1) False-color image. (2) Ground truth image.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIn the false-color image: (a) \u003c/strong\u003e\u003cem\u003ePaecilomyces lilacinus\u003c/em\u003e\u003cstrong\u003e, (b) \u003c/strong\u003e\u003cem\u003eAspergillus niger\u003c/em\u003e\u003cstrong\u003e, (c) \u003c/strong\u003e\u003cem\u003eAlternaria alternata\u003c/em\u003e\u003cstrong\u003e, (d) \u003c/strong\u003e\u003cem\u003ePenicillium citrus\u003c/em\u003e\u003cstrong\u003e, (e)\u003c/strong\u003e\u003cem\u003e\u003cstrong\u003eTrichoderma longibrachiatum\u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e, (f) \u003c/strong\u003e\u003cem\u003eCladosporium cladosporioides\u003c/em\u003e\u003cstrong\u003e.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6292446/v1/7e93448bd79859af19d920b0.png"},{"id":81023098,"identity":"8a14cfa7-075c-4896-a11d-702d433af603","added_by":"auto","created_at":"2025-04-21 10:07:21","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1000859,"visible":true,"origin":"","legend":"\u003cp\u003eMold stain image acquisition system\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6292446/v1/165b9ae73676b908d55d502e.png"},{"id":81025242,"identity":"a5ae30dd-837c-4296-a49c-cf5b11f60c6d","added_by":"auto","created_at":"2025-04-21 10:31:21","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":252721,"visible":true,"origin":"","legend":"\u003cp\u003eTPMFN network architecture, including three data fusion methods\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6292446/v1/bcf688ca3db42bd0e753367b.png"},{"id":81023096,"identity":"e643e497-0c92-42ab-b469-1f5a2d64c18e","added_by":"auto","created_at":"2025-04-21 10:07:21","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":88565,"visible":true,"origin":"","legend":"\u003cp\u003eThe network structure diagram of the spatial feature extraction model\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6292446/v1/05b3bbec75505bb907ee55c3.png"},{"id":81024627,"identity":"eeefc95d-f752-46b8-83f8-0f4d7a90b161","added_by":"auto","created_at":"2025-04-21 10:23:21","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":99719,"visible":true,"origin":"","legend":"\u003cp\u003eNetwork structure diagram of the spectral feature extraction model\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6292446/v1/3c31b004802ac69436956298.png"},{"id":81023102,"identity":"56ba6ba4-eb92-4a15-9c4f-94367529a170","added_by":"auto","created_at":"2025-04-21 10:07:21","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":164471,"visible":true,"origin":"","legend":"\u003cp\u003eThe network structure of the spatial-spectral feature extraction model\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6292446/v1/1904f8d5e24226ceb1b5326d.png"},{"id":81023107,"identity":"4d1c06f8-18f7-4d88-8327-39414b9809a9","added_by":"auto","created_at":"2025-04-21 10:07:21","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":144327,"visible":true,"origin":"","legend":"\u003cp\u003eComparative analysis of feature fusion methods\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-6292446/v1/576a41e52f6afaa5a64d7498.png"},{"id":81024634,"identity":"d77e7608-9581-4669-a465-fc08f9800098","added_by":"auto","created_at":"2025-04-21 10:23:21","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":173098,"visible":true,"origin":"","legend":"\u003cp\u003eComparison chart of overall accuracy with different proportions of training samples\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-6292446/v1/ca2ec633253617216f16c7f5.png"},{"id":81024343,"identity":"ee40b6a3-89d0-40e7-bab3-9666ef7693db","added_by":"auto","created_at":"2025-04-21 10:15:21","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":814406,"visible":true,"origin":"","legend":"\u003cp\u003eClassification results. (a) Ground truth. (b) SVM(OA=61.18%). (c) 1-D-CNN(OA=70.90%). (d) 2-D-CNN(OA=85.09%).(e) Spectralformer (OA=91.71%).(f) SSFTT(OA=94.39%).(g) HybirdSN (OA=95.38%).(h) TPMFN (OA=98.52%)\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-6292446/v1/82aaa830324d439e20f836b9.png"},{"id":94490646,"identity":"e5d1328b-61af-4b00-846b-dabd54bf582f","added_by":"auto","created_at":"2025-10-27 17:13:24","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4819769,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6292446/v1/d573e3c3-7e14-4f48-bf16-be87115dd25a.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Mold Spot Detection for Paper Artifacts Based on Multimodal Feature Fusion","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eAs pivotal witnesses to the splendor of Chinese civilization, paper-based cultural heritage artifacts have profoundly shaped intercultural dialogue and promoted the evolution of global civilizations. From classical paintings and canonical texts to epistolary documents and archival records, these artifacts offer invaluable research resources through their substantive content and refined sheet-forming techniques\u003csup\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/sup\u003e. Nevertheless, the inherent hygroscopic fragility of paper matrices renders them susceptible to biodegradative and environmental stressors during storage and display processes, thereby escalating the complexity of preventive conservation interventions.\u003c/p\u003e \u003cp\u003ePaper is one of the most vulnerable materials to mold growth due to its high content of cellulose, hemicellulose, and other nutrients. Mold secretes cellulase to degrade and absorb these components. During this process, the structure of the paper gradually deteriorates, and the bonds between fibers weaken, leading to a decline in its mechanical strength\u003csup\u003e[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/sup\u003e. Furthermore, this biodegradation process may be accompanied by a series of complex chemical reactions, which further accelerate the aging and embrittlement of the paper, thereby severely damaging the historical and artistic value of cultural relics\u003csup\u003e[\u003cspan additionalcitationids=\"CR4\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/sup\u003e. Current mold remediation approaches for cultural relics are primarily categorized into three modalities: mechanical, physical, and biochemical methods\u003csup\u003e[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e. Mechanical interventions employ tools such as soft brushes, conservation scalpels, and HEPA-filtered vacuum systems to remove microbial colonization, though complete microbial elimination remains unattainable\u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/sup\u003e. Physical techniques including UV irradiation and laser ablation demonstrate efficacy in microbial decontamination, yet show limited effectiveness against most bacterial strains and dematiaceous fungi\u003csup\u003e[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/sup\u003e. While biochemical approaches prove more potent, the taxonomic diversity of molds necessitates species-level identification to devise targeted conservation strategies that prevent secondary damage, as monotherapeutic interventions often fail to achieve comprehensive eradication\u003csup\u003e[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/sup\u003e.Therefore, the precise identification of fungal species serves not only as a prerequisite for mold remediation, but more crucially as a pivotal measure in safeguarding paper heritage artifacts against prolonged fungal colonization. Current conventional detection methodologies for paper-based cultural relics encompass morphological characterization, molecular biology techniques\u003csup\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]\u003c/sup\u003e, and biochemical assays\u003csup\u003e[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/sup\u003e. However, these approaches rely solely on singular technological modalities - be they image-based or spectroscopic - which impedes comprehensive mycological identification and consequently manifests inherent limitations.\u003c/p\u003e \u003cp\u003eHyperspectral imaging technology, as an advanced detection means fusing image and spectral information, has shown a wide range of application prospects in the field of mold detection with its high sensitivity, multi-dimensional information acquisition capability and non-destructive detection characteristics. At present, this technology is mainly applied in the fields of food quality assessment\u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/sup\u003e, crop disease detection \u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/sup\u003e and preliminary progress has been made for the identification and assessment of mold contamination in the field of cultural relics protection.For example, Lu\u003csup\u003e[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/sup\u003eet al. proposed an automated labeling method based on hyperspectral images for the detection of mold damage in murals, which effectively addressed the issues of time consumption, subjectivity, and inconsistency in traditional manual labeling methods. By integrating the spatial and spectral information of hyperspectral images, accurate identification and labeling of mold on mural surfaces were achieved, demonstrating high practical value and research significance. Dai\u003csup\u003e[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/sup\u003eet al. systematically studied the spectral characteristics of simulated samples of paper artifacts affected by foxing based on hyperspectral imaging technology. By applying band arithmetic and the minimum noise fraction (MNF) method, the ability to extract and differentiate features of the infected areas was effectively enhanced, and a discriminative model based on the K-nearest neighbor method and BP neural network was constructed, achieving an outstanding accuracy of over 79% in foxing detection.The above research provides important technical support for the accurate detection and scientific control of mold contamination and further extends the application potential of hyperspectral imaging technology in cultural heritage conservation.\u003c/p\u003e \u003cp\u003eIn this study, we integrate digital image data with hyperspectral imaging data, employing multimodal data fusion technology and machine learning methodologies to achieve efficient detection of microbial colonization on paper-based cultural heritage. Through evaluating the impact of different feature extraction methods on mold identification and comparing their accuracy and robustness, we establish an optimal detection model. Furthermore, we extend the applicability of multimodal data fusion technology in cultural heritage conservation. This approach not only provides an effective solution for early-stage detection of biodeterioration in paper artifacts, but also pioneers new methodologies for multimodal data fusion in heritage science, demonstrating substantial scientific and practical significance.\u003c/p\u003e"},{"header":"2. Experiment section","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003e2.1 Simulated mold infestation samples\u003c/h2\u003e\n \u003cp\u003eTo ensure the diversity and accuracy of the experimental data, this study selected cotton-linen raw Xuan paper, cut into 5\u0026times;5 cm samples, as the base material for simulating mold contamination. Based on literature from the past five years on common mold species found in the preservation environments of paper-based cultural relics, six representative mold species from different genera were chosen as experimental subjects, including \u003cem\u003eAspergillus niger\u003c/em\u003e, \u003cem\u003ePenicillium citrus\u003c/em\u003e, \u003cem\u003eTrichoderma longibrachiatum\u003c/em\u003e, \u003cem\u003eAlternaria alternata\u003c/em\u003e, \u003cem\u003ePaecilomyces lilacinus\u003c/em\u003e and \u003cem\u003eCladosporium cladosporioides\u003c/em\u003e. These six mold species belong to distinct genera, reflecting both the biodiversity and ecological adaptability of molds. Widely distributed in natural environments, they can potentially cause deterioration in cultural relics. Therefore, it is essential to prioritize monitoring these genera during mold detection and conservation processes. The paper substrates and fungal strains used in this experiment were supplied by the Key Scientific Research Base for Cultural Relic Pest Control under the State Administration of Cultural Heritage. A photograph of the artificially inoculated mold samples is presented in Fig. 1.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003e2.2 Mold stain acquisition system for paper-based cultural relics\u003c/h2\u003e\n \u003cp\u003eIn this study, a hyperspectral image acquisition system for mold stains on paper-based cultural relics was built, as shown in Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. The system consists of an iSpecHyper-VS1000 portable hyperspectral imager from Lyson Optical, two halogen light sources with adjustable brightness and color temperature, and a Canon RF 24mm focal length lens. The specific parameters of the hyperspectral imager are listed in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. After multiple imaging tests, the optimal parameters were determined as follows: a focal length of 4615mm, a frame rate of 16, and an integration time of 56338. Images captured with these parameters exhibited excellent quality, with sharp edges and minimal distortion.\u003c/p\u003e\n \u003ctable id=\"Tab1\" border=\"1\" class=\"fr-table-selection-hover\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eiSpecHyper-VS1000 hyperspectral imager parameters\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMain Technical Specifications\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSpecification Parameters\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSpectral Range\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e400\u0026thinsp;~\u0026thinsp;1000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSpectral Resolution\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026lt;3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eF-number\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eF/2.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eImaging method\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePushbroom Imaging\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNumber of spectral channels\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e300\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDetector\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCOMS/InGaAs(TE Cooled)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAD Dynamic Range\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFrame Rate of Spectral Camera\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e50\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSpatial Resolution of Image (Pixels)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1920\u0026times;1200\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDetector Pixel Size\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5.86*5.86\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\n \u003ch2\u003e2.3 Experimental setup\u003c/h2\u003e\n\u003c/div\u003e\n\u003ch3\u003e1) Evaluation metrics\u003c/h3\u003e\n\u003cp\u003eIn order to verify the validity of the models, this paper quantitatively analyzes the classification results of each model using three commonly used classification performance assessment metrics: Overall Accuracy (\u003cem\u003eOA\u003c/em\u003e), Average Accuracy (\u003cem\u003eAA\u003c/em\u003e), and Kappa Coefficient (\u003cem\u003eKappa\u003c/em\u003e). Among them, OA reflects the overall classification accuracy of the model across all categories, AA indicates the mean classification accuracy for each category, and the Kappa coefficient measures the consistency between the classification results and random classification results, thus providing a more comprehensive evaluation of the model\u0026apos;s performance. Meanwhile, to further visualize and compare the classification performance of each model, this paper also conducts qualitative analysis by visualizing classification maps to observe the spatial distribution characteristics and the accuracy of the classification boundaries, in order to comprehensively evaluate the classification performance of the models from both quantitative and qualitative dimensions.\u003c/p\u003e\n\u003ch3\u003e2) Comparative analysis of models\u003c/h3\u003e\n\u003cp\u003eIn this paper, several state-of-the-art models are selected for comparative analysis with the proposed model, including: SVM, 1-D-CNN\u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/sup\u003e, 2-D-CNN\u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/sup\u003e, SSFTT\u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/sup\u003e, SpectralFormer\u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/sup\u003e, and HybridSN\u003csup\u003e[\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/sup\u003e. To ensure the uniformity and validity of the experimental data, a unified preprocessing approach is applied, where the Savitzky-Golay(SG) smoothing filter is used to remove noise, and principal component analysis(PCA) is employed to reduce the number of spectral bands to 30 dimensions. This helps eliminate redundant information and highly similar features, thereby improving recognition accuracy.\u003c/p\u003e\n\u003ch3\u003e3) Experimental setup\u003c/h3\u003e\n\u003cp\u003eIn order to validate the multimodal feature fusion-based classification algorithm for hyperspectral mold detection proposed in this paper, all experiments were conducted using Python 3.12 and implemented in the PyCharm 2024.3.1 integrated development environment. The experiments were performed on a high-performance personal computer equipped with an Intel(R) Core i5-14600KF processor, an NVIDIA GeForce RTX 4070 GPU, and 32 GB of RAM.\u003c/p\u003e\n\u003cp\u003eDuring the experiments, all deep learning models employed Cross-Entropy Loss as the optimization objective, in conjunction with the Adam optimizer for parameter updates. Stochastic Gradient Descent (SGD) was used for model training optimization. The learning rate was uniformly set to 0.001, and the number of training epochs was fixed at 50 to ensure model stability and convergence, thereby enabling a reliable evaluation of each model\u0026rsquo;s performance.\u003c/p\u003e"},{"header":"3. Related Work","content":"\u003cp\u003eIn this section, we first introduce the TPMFN model proposed in this paper, followed by a detailed explanation of the design logic for feature extraction along the three different paths within the model and their roles in multimodal feature extraction and fusion. Finally, we summarize the advantages and potential of the TPMFN model. Here, after the data block division, the input hyperspectral image features are denoted as , where B represents the batch size, H and W represent the spatial dimensions, and C represents the number of channels.\u003c/p\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n \u003ch2\u003e3.1 TPMFN\u003c/h2\u003e\n \u003cp\u003eThe model aims to enhance the accuracy and robustness of mold stain detection through the efficient fusion of multimodal features. The core design concept of the TPMFN model is to fully exploit the characteristics of three different modalities for feature extraction: the spectral dimension of hyperspectral data, the joint spatial-spectral dimension, and the spatial dimension of RGB images. By employing a deep fusion mechanism, the model achieves a comprehensive representation of global features. The network structure diagram is shown in Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\n \u003cp\u003eIn the TPMFN model, the input hyperspectral image is first processed to extract three RGB bands, generating an RGB digital image. Then, different feature extraction methods are applied to different modalities.For the RGB digital modality, 2D convolution neural network(CNN) is employed to capture spatial characteristics such as edge morphology and color variations in the pseudo-color image. Subsequently, a spatial attention module is introduced to enhance the spatial texture representation of the mold stain region while filtering out interference from non-mold stain areas.For the hyperspectral data modality, feature extraction follows two paths. In spectral feature extraction, a Spectral Transformer is first used to capture subtle spectral differences of mold stains across hyperspectral bands. Then, Spectral Attention is introduced to emphasize key spectral bands, generating spectral features.For spatial-spectral feature extraction, a Hybrid Convolutional Network is employed to jointly analyze the spatial distribution and spectral variations of mold stains in hyperspectral data.\u003c/p\u003e\n\u003c/div\u003e\n\u003ch3\u003e1) Spatial features\u003c/h3\u003e\n\u003cp\u003eThe 2D CNN is a deep learning model specialized in processing 2D data. Its core mechanism utilizes convolutional operations for local feature extraction, where parameter-shared kernels capture low-level features (e.g., edges, textures, structures) within localized spatial regions, while progressively learning higher-level semantic patterns through layered hierarchies. In this work, we first extract and flatten the three RGB bands from input X. A three-stage 2D convolutional network then hierarchically extracts spatial features from the hyperspectral image, with optimization applied to features at each layer to amplify discriminative spatial regions. The flattened outputs are subsequently fed into a three-layer fully connected network for progressive dimensionality reduction, ultimately generating category predictions. The spatial attention mechanism enhances critical regions by computing spatial weight distributions, with the attention energy defined in Eq.\u0026nbsp;(\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e):\u003c/p\u003e\n\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\n \u003cdiv class=\"EquationNumber\"\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"525\" height=\"41\"\u003e\u003c/div\u003e\n\u003c/div\u003e\n\u003cp\u003eIn Eq. (\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e), Q is the query matrix, and K is the key matrix.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eThen, a Softmax operation is applied to each row of E to obtain the attention weight A. The formula for calculating the attention weight is defined in Eq.\u0026nbsp;(2)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"646\" height=\"109\"\u003e\u003c/p\u003e\n\u003cp\u003eIn Eq. (2), \u003cem\u003eA\u003c/em\u003e\u003csub\u003e\u003cem\u003eij\u003c/em\u003e\u003c/sub\u003e represents the weight distribution of the pixel spatial position at the i-th row and j-th column.\u003c/p\u003e\n\u003cp\u003eThe specific network structure diagram of the spatial feature extraction algorithm is shown in Fig.\u0026nbsp;4.\u003c/p\u003e\n\u003cp\u003eFigure\u0026nbsp;4. The network structure diagram of the spatial feature extraction model\u003c/p\u003e\n\u003ch3\u003e2) Spectral features\u003c/h3\u003e\n\u003cp\u003eSpectralFormer is an improved Transformer-based model that enables local spectral representations from multiple neighboring bands at each encoding location. This is achieved through the Group Smart Spectral Embedding (GSE), which enhances the capture of fine spectral variations, and the Cross-Layer Adaptive Fusion (CAF), which improves the transfer of information between layers. In addition, SpectralFormer incorporates a cross-layer skip connection, which adaptively learns to fuse their \u0026quot;soft\u0026quot; residuals, gradually passing memory-like components from shallow to deep layers.Although SpectralFormer is good at capturing global sequence information, it is less effective at modeling the local contextual information of hyperspectral data. To address this limitation, this paper proposes an improved spectral feature extraction model based on SpectralFormer.\u003c/p\u003e\n\u003cp\u003eIts structure is shown in Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e. First, the spectral attention mechanism is introduced, performing global average pooling to obtain the global representation of spectral channels. It then assigns weights to spectral bands to emphasize important ones, followed by applying linear projection for spectral embedding. Classification tokens and positional embeddings are then added, followed by sequence modeling using the Transformer module, and finally, the classification results are output.The input is the extracted spectral information\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"129\" height=\"38\"\u003e\u003c/p\u003e\n\u003cp\u003ewhere N represents the number of features. The spectral attention process is formulated as shown in Eq. (\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e):\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\n \u003cdiv class=\"EquationNumber\"\u003e\u003cimg src=\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAA1wAAACXCAYAAAABIlouAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsMAAA7DAcdvqGQAAE/lSURBVHhe7Z1ptDxFecZlETTJhyj7ySf2JSdfZFX4ogiCSmKisiibAgFcMCIQowYxQJTdBIMCEiOLCjkGEAGTuLBooqzGHGRTlEUQlwiSqKwTfsU81/df/6qe7lnunZn7/M7pc+9UdW1vVXe/z1R1zfN6xhhjjDHGLDLPM2a50B/zxhhjjDHGLBp9V9SY+ac/5o0xxhhjjFk0+q6oMfNPf8wbY4wxxhizaPRdUWPmn/6YN8YYY4wxZtHou6LGzD/9MW+MMcYYY4wxZsxYcBljjDHGGGPMhLDgMsYYY4wxxpgJYcFljDHGGGOMMRPCgssYY4wxxiw6/e0EjJlr0lhPI94YY2acBx98sLfJJpvwAC8e++23X+/pp5/un22MMWapefbebMzck8Z6GvHGGDPjfPGLX+y95z3v6T366KO9W265pbfFFlv0brrppt4zzzzTO+qoo3pf+MIX0v/GGGOmg74/asxck8Z6GvHGGDPjXHLJJb0f/ehH6f/TTjut98pXvjKJLzjvvPN6v/zlL9P/xhhjpoO+P2rMXJPGehrxS8y5557bW2uttXrHHXdcP8QYY4bjscce6+2yyy69U045xUsIjTFmiun7o8bMNWmspxG/hPDexcYbb5zesTj11FP7oWYS4Iiy5GrttddO9kbkfuYzn5n7ZVZ33XVX7+ijj+7dfPPN/ZDlTbTHPPb9rbfeusJywqWEa+7666/vHX/88b3dd9+99+53v7uTCKQt9NUdd9yxpG0pteOpp57qx3aDNh1zzDFL3iZjlhNXX3117+STT+49+eST/ZD2fOMb30hfYD388MNjv2aTN2rMnJPGehrxQ4BQ2mijjZLjHg+c+OjYIqKazuFdCwRAnm4piOKvVl9gudIqq6zSeM60gcPEEqv9998/La2S3ak7N9J5dXw+97nP9TbffPMFYRnbnfffnnvuuVK/6pCNDjjggJXOmSX75fbA+V1nnXVWaA9HW3sceOCBK53DQ30p7cG7XHE54VLB/WTTTTddwT6XXXZZZ9vQZwjIiy66aEnsSjs222yzkdsRufjii3tbbrll78ILL/QspDET5ogjjuidddZZK31Jcs899/Te/OY391ZbbbV0XePTfexjHyt+mfKd73yn99rXvrZ34403jvWafbZcY+aeNNbTiB8SnHiWAZIXB0sDS8jJ3XbbbXtXXnllSieuuOKK9CBHCCylkyao24c+9KEF5+Kcc87px6xIbBMOXmzTNIJIpL5RFGJ7wnC+jzzyyKkWjMOAo5q3GegrCSeERZzli/1PHLMkORLcBx98cHJGZwXZI28TbZZwos3RsS/ZI79Osceqq646NfagLdxPpsWRR9Suu+66yZm5//77+6HdUB5LJbogtuO+++7rhw4P+a233noWXcZMEMTWRz7ykZVmthBQXH9bb711b9ddd033d+7z+D0nnHBCcSZMaW644YaxXbPPlmfM3JPGehrxI4BDxrfJ5FdaEkg835TXlgvK6Z8mZz+2qTR7Qfxee+2V2rRUzk8X1B5uqCU74yRvt912cyW4JIhrO9Mhknm4vOIVr1hpM4WmOEBgzErfC+zBTNbll19etQeiiTbnM0NNcTBN9mAsszV8rZ1LAfbjG+RRt6VXPqPOLg2Lyt93332HXk6Ys9RtMmaeOfbYY9Ps9G9+85t+yG/ZY489el/72tcW7kk///nPe+985zsXZrseeOCB4jXJ/Z482aBoHNcszqgx804a62nEjwjOFvntvPPOKzioOPrMgOHs1SDttMxuRTSLUWoT3/g3tWnawAllqWRNcDHbU4ubReR0530XUf/iBMexp7SM5zwO6Hfeg5u28dpEtEdtmV20RxQFuT1ywcAsBfYY17edo0I7+PIAZ2Fa0OzhqKKCew/fRHMtDztTNgpvectbkvAepziKbRrHrJkx5jl472qNNdbonX/++St9QUIc72Tm921E1wte8IJ0v//yl79cvK9zzvrrr9/78z//894TTzzRDx2eZ8syZu5JYz2N+BHRbELutLNMjRmsJjhnGsVLrU04l4PaNG00CS452vMkuNSmpt9d0pLCfAaTtDvuuGNRcGHHvffee2Hr8VlB9mia9ZEoqNmDuJIY22effaZC3ERhyMF4HtemGaeffnoaL8OISurFe1yjLCeMaEaI9+S61kftGGZ2inbwrfa4lhNG1KaTTjppKBsbY1aGGSy+IKnNVNXgfj8oHcsUV1999d5tt9028j0WZ9SYeSeN9TTiR0RL1sgTJxfaiK1pptSmSYot3m3jm3nK03thvBNXEqPULe42yLHbbrv17rzzzv4Zz4GTVNoEhIMySFOK08wQdeJbLC35ZHc7PuN8cx7v7GgG6aqrrlqoP84u776VbsSlunM+u56RV7R7PCQEJIQVni/7k+M9SECWBBd5UzdmXEvig7il+vFc2Z4NL6JdSofEBkR7lN5HEyXBpdmrmj2Wy48JjyK4RhFIJUYRcKMIrkmKokmKOWOWI5rd4ucxHn/88X7oYDTDxS6iTbsZ8puH5H/IIYeMPMv17DPLmLknjfU04scAThl5/tmf/dnMiy2Bc7IYbSJfRARODURRkosGCQ7sLaGBMJIIyZ1ikPAqiZDaTB5tJ4w8+Yv4++QnP7mSKPrTP/3T5Mh99rOfTeUiDKJQjXWhHogyHCuVRV6vf/3rk0MfZ5TY4IEw8slnmrATdf7gBz+4gtgC4khXe/8KVH/OUx0JQ9Awe1USY+RLn8R6LAYaC7JFmyO+a0W9m96/AsrgwUwZmgVra49Jz0jEupXa2nWb9WEYRXAhZBmr7OzF9aMdIQ866KBqfzSBPViChz26Lu0bRXCxnLDWjkceeaR/1nCM0iZjzMowA8UXJIOEUw7peEaX3vmKIOjWXHPN3u///u/3fv3rX/dDh+PZ+4gxc08a62nEjwFEg5wizQiNE70n1uUYtR6TbpPACcrfY8MJYWOOKIIIQyiU3k2KMz95XYcRXNCUp2xTSqclbLlQqoWrHIRYXK6n81W+0uDo1zZpqJURwR5afqZ8ERDkiV1zMcb5S7GUMIoNZgDYTZG6COzAFwG1dkK0R00wRHtIcEV75GJM9liMpYTMspW2redoatM4GVZwYSdmo5hNZjYY4RvrjwgeRqzoXaqum1cMK7hoB+Nv3O2IDNsmY8zKbLXVVumefeKJJ7YSXPxG1w477JCuwTPPPLPVNfjCF74wXf+1d73a8mwexsw9aaynET8GeChr+VrunM8qi9EmiaiScOGbZMSIkPNcmsUCnCnic0GmdnQVXG3S5SIJcNZL9SCcNMxMxfo31UFtUjnYoCa2ZEvObyO4KI9ldnxmNoe65nGAsImCb7FQ22lLtKNQXWvjIQq2NoKrjT0oZzHtwbJFxk1e1t13393/b/IMK7i0jTq2Yzt39SGzt3z7TDgzRl3tKHHSVegMK7hiO9jCXe3gt7RGaUdk2DYZY1ZGG1985StfGXjf0rtenK+D5wb3/qZrWoKrtoV8W57Nw5i5J431NOLHAN+08+AlXxzieQDHUm2qObVtUD613/SSY805epcpR2KCutTEX23WiRvnNAiuHNp0xhlnpHxoV6mc2G6+Zcc+tX6Igqupv2LdmaWhf1RuLjCwKfGjOJMlyJM61N51w4ZyQActBVQ7chQ/yB6aRcrtwfmTsgdCivafffbZnfNid62mcTVuhhVcpKMP82Vy9AtL6Gj/MEJF+S6W4CIdwmrc7YioDAsuY0aD38piuR/PzDaCC3h3iy9QeEeY5wVpB72ftcEGG6RzLbiMGUwa62nEjwhii3d4EFrkO8jRngVwKtWmNuKhhkRLk104RxtO6GDpTtwEI+ZTE1w18aS0TcKpFNcm3TCCSxtAkB4BpXfQSuWAyqLtTcKhreBS/ag7/8cZM8qS+KAuk1hKiF0RMbSnJKiiUBr0e1La7KJ0XlvBJXFHmxFU0R5RjCG4xmGPQe1vYpJiS0KGerU9agJBYgS75ZtbKI5ra5idC1XPDTfcsJh+Eu2grvmGFk1xXZHgok3eOMOY4dGGGVzXbQVXhNUQXIuk/9WvftUPXRnNjFlwGTOYNNbTiB8BiS1oct5nCYktGEebyI/0tRkuwHmJsz0cscwouGrOc62u0yS4tCwyzuQNsjHx2qqdo7acDRt2EVy8k6KNIUSM4xu+SS2dY0zQ5tIMlwQJtm16T4r2IqiwG2IoR/Ft7MGDc5tttkltzu0R48Zlj2FmuBZ7ZksgBrrOcGkZXmkpJ/3Lu13DzuZIUHVNr3Z0meGK7cjT0Q5mnYdtR0SCaxx5GbOcGWaGK3LPPfcspG8SXNo+3oLLmMGksZ5G/JCwc52ECchBJ+/aLMywaPasyzFMHRazTSVwkikfZ5QyJVokJgjD+Sk5qTURpDYsteCqbWgxqA6aVVH6UpkQBVfNRqB81KfxPNW9yc6TBnswq4T9mmZ/JIZwUkvnRcFFW2oP3miPfKZMZcgeXR/e44JZUdqzFAwjuCSKSrvuYdPSEr22LKbgkhCaRDsiKseCy5jR0Ttcl1566VD3bImpJsHld7iMaU8a62nEDwHChCNH28Pzd9aotQknRW0a1bEo8aEPfWglAYGDidiJ4mKQ4NA7XLlQmAbBFQVjLnJqdSBN3Kkx5lGaPYOaqIs0nRNtvBi78JWQ4GqqQxRTTcsOY1sHCa7SOdNgD2ZZGIsRwq699tpqu8eJhEoXx4Wlnlw7teWEjN9hhcVi7lJIWbSjtpxwlHZEvEuhMeOj6y6FOQiuPffcs/E3vLxLoTHtSWM9jfgOIARYWoRzX0JOf00UTCN6p2ip2oQTFN+bAYmLWKYEEH1XWiKGc4xwGZeoWmrBxbK73C7RBiXBpLL5prwkyAB7l+wExJE+r2MJysJJpC61WaZhaCOmJISaNhEB1bGpfvqtKO1CGGl6RyyCANpiiy2KeYxCaWaLz4ceeuiiCcCugosxypLBkoAlL5aL3n777Qt2wnZbbrllq40naDtChz7pOrPUVXDRDpYM0o48TakdwHhjpkrXRBsxNkqbjDErc+yxx/ZWX331xt/hYulh6Z7GO2Drrbde74Ybbqje8+KyxaZZsDY8m4cxc08a62nEt+Chhx5KD1mcYn4IOHeCBMvhyJuDGaHaedOA2oSzSZtqDjptwhlQm2rnDYuc/PhDviozF1YSJ9Qlns9W04SXlj2yvbzqn78zRP9ICPE7T3w+/vjjU76x3cTFdHL4GQ/XXHNNP/Q5mLFTXBQ1aiciTRuCUG+9n6XzmWVkZitPL5QPaXL74CTiCNbSqr21d5rIu81MJv2AIOQ8lVkSgMNC/sxy0Q4Ev/JFgGy//fYpvLbDYSTao+k9r5o9EFzYo0lsSNSpjHHZACGCDejn/BinwB2EhEpbwYVNsRsb4dxxxx3JHtxreG8thoEECrZrI7gk5riGum640VVw0Q5mnqizhFVsRy626C/iyL9JrOXoXNrkDTOMGR0Jope85CXFHzHWDyNzzV1wwQUL1+gll1ySdipsElvA73axMcegnQzb8Oz93Ji5J431NOIHgMPF+fHgQuVBKaIQyA/STxs4Hnk9aVOcrRnUpkHOUVvOO++85MhoIwXyxwHBGSuhGTmJDo58V0OgfzQblB/RwcahV7nkQzqJsHgwY0UZpTxxrBA4JXtJiODA0UbVm40paKNEkMK4ice2abYpnhePfDZLYrAkIshDgjKHuHwTjRps4MD5gjKp2zhFQKmfGReUVap/jcWwB2MGITguwdUktjhOPvnksZTThq6CS/BlgtqAoGK2tjQ+sB0/OtrGdhJotL9rfboKLsF20WycMagd+aYmlMc1MWiWS2066aSTOrfJGFNGuwiyEiC/rzCL9dKXvnRhhQYHX7z9wz/8Q+/hhx8eeB/SDNqoywnh2bKNmXvSWE8j3pg5AgeWWZ3ae16TAKcRZ3axREAXoj0mNSs0bsG1nOgiuPRO1TDbyS82XBPM9A0SeLX3xIwxw6Pt4c8///zOX7IMgt/g4su4UWe3oO+PGjPXpLGeRrwxc4ZmJ2tL5cYNYqvNe19LhZYoTsoeFlzD01ZwaSZoVt5zQkgNevdv1tpkzCzB6gZWRZSWFQ6Llh2y8mEc12zfHzVmrkljPY14Y+YQnDlE16SFEGJG73NNM9gD0TXICR4GC67haSO4WGLJsr5hlhIuBfF9rhqcw8v5s9ImY2YR3tc6/PDDxzIbxW90scHPoHe8utD3R42Za9JYTyPemDlFomtSMztd3nOaBiS6xm0PC67hGSS4eBeMHSAvuuiimbAt18Rhhx3WuIsk74XhuF144YUWW8ZMGN654h3JUX4vC7HFzzZwnxrnNdv3R42Za9JYTyPemDmGzSeOPvro4q6Fo8LM1iTynSTRHuNy4C24hqdJcDELRF/FnQ2nHerbNA5oE9tVz1KbjJl1eKfr7LPPHup9LnY9ZFfcNhtqdKXvjxoz16Sxnka8MaYzrI+PW7YzexQ/LycsuIZn0AzXLMHOhHEJ73K+Jowxg+n7o8bMNWmspxFvjOkEm2TE7do5+FmBxfox3mmCGYu4hfsk3hGbV7CdtlznmOXNI9gkI24zzbHhhht690FjTJVn7xPGzD1prKcRb4wxxhhjzCLS90eNmWvSWE8j3hhjjDHGmEWk748aM9eksZ5GvDHGGGOMMYtI3x81Zq5JYz2NeGOMMcYYYxaRvj9qzFyTxnoa8cYYY4wxxiwifX/UmLkmjfU04o0xxhhjjFlE+v6oMXNNGutpxBtjjDHGGLOI9P1RY+aaNNbTiDdmjuiPb2OMMcYYY5aU5JsmD9WYOaI/vo0xxhhjjFlSkm+aPFRj5oj++DbGGGOMMWZJSb5p8lCNmSP649sYY4wxxpglJfmmyUM1Zo7oj29jjDHGGGOWlOSbJg/VmDmiP76NMcYYY4xZUpJvmjxUY+aI/vg2xhhjjDFmSUm+afJQjZkj+uPbGGOMMcaYJSX5pslDNWaO6I9vY4wxxhhjlpTkmyYP1RhjjDHGGGPM2LHgMsYYY4wxxpgJYcFljDHGGGOMMRPCgssYY4wxxhhjJoQFlzHGGGOMMcZMCAsuY4wxxpg549Zbb+2tt956vY022qh333339Z555pl+jDHj49vf/nZvgw02SOPshz/8ocdZBQsuY4wxxpg55Morr+wdeOCBvaeeeqof0o6rr766t8MOO/RWWWWV3lprrdW78MILe1dccUU6nn766f5ZZtagX1/60pf2Vl111d6LX/zi3gUXXDCWfmWcveUtb+k98cQT/RCTY8FljDHGGDOHnH766b0vfOELrZ3pxx57rPeqV72qt+222/ZuvPHGNFtB2PHHH99bZ511ejfccINnMGYQ+nC33XZL/ao+VL+uu+66vW9+85sj9etHP/rRTuNsOWLBZYwxxhgzZzz00ENJPN1///2tnGmJrZ133rn3i1/8oh/6HOSFw+6libOHxBb9+j//8z/90OegX3fffffevffeO3S/jiOP5YAFlzHGGGPMjINjffTRR6eZqG222aZ39tln9w444IC0nPDuu+/u/c3f/M0KSws5/y//8i8XBNkZZ5yRZjtKs1ic+4//+I+ewZhB6Ffe5fvWt75V7NdPfepTC2OEGa8nn3yyH/tc/Hvf+94VRDthxxxzTBorjLNPfOITadmqlxM2Y8G1hPCi4Yte9KLe4Ycf3g+Zbn72s5/13ve+96V6DwMX5Wc/+9n+p8WDep900km9jTfeOP015nOf+1z6xo9vc5fiG7lhymccsxyEB9xPf/rTfmgzpDn55JN7m2yySe8jH/lIq7KGSbPUfP3rX+/tvffe6X2T3CHE1ryvcNFFF81EW8bBf/3Xf/XWXnvt3mGHHWYHeZmAE8z95N3vfnfvkUce6T344IO9XXfdtXf55ZcvjAHes3nPe96zILoQZ1oGpvT77bffCg63mW3oV5419GsbQcQYOeqooxbGAMIqjiHlxzhjFpRxxufLLrvM95oBtBJcz3ve86rH9ddf3z/LdAXnX3acdhBZ3Iy///3v90N+C0IKMaO27LXXXsXzAMFFPE7dYsD4pN6qmwXX8oZx+f73vz+9LMx46CJ4xkFePg5R2/IZywgH0l177bX90Dqcz4MQEUKaNuJpmDRLzdve9rYkDqkvR/7QZ6kL7cHWy8UhOOWUUxbGStfNEqYZdt3jW3X1tQ42ddhzzz0Xxm1+8AUCfc9L/bKLDp4JizkucFgZi7W66th00017Bx98cO+OO+5odQ3SNvpdbaGcffbZZ+E9LME7Xccdd1zvyCOPXOGdG5aFbbbZZslWszRmGBPrr7/+SvZjTLzxjW9cqb91cG+jnW9961tXOkdxS4EETa3eOhgfBx10UO/2229vHB/06xZbbJHGeds2MSOmMRLFFlAm40x5Ud83v/nNaVZ0Ma+jWaS1p8+DmNkYdTafTTtqM1izMsNFPRFUJZFE3WkDjivfvGt8EFYTVQg0zl0s0QWqmwXXdIGzPGmHvlQGTgUOz2ILLqHyuwgurpeuM1yw3XbbpbHfRTwNk2YpueqqqxYclPyh7xmu+XKCcPBYvkSfavc8tZE4ls0pLhcbgOAgHsfxRz/6UT908aGuEoCveMUrVnhniqVdhx56aIrTc6upH5mV2GWXXVbIoxQmKHf//fdfwQHXFvJdHPNpAVtKONHv7LwXxwTL5FZbbbUUV1pWh8BYffXVUx4PPPDAVNwn1Cbq/fKXv3yFd68YH1zbxOk+XeszCdKuIpLrgzESZ8UYUzyzYl1KYaZMp6kVzRTwwDft4BttbDarUH/EU2kpIBcagisKpyjMWX5Yg7HETNdiobFrwTU9MLYQHZN8uGnpXl7GLAquYdGMVRfxpFmhWRFcLCmUczpvAsOszBe/+MXkbCJUWD4XaYoDBNepp546FeOEuuLo54JLxBm5Sy+9tFpnzoszETjZbOnOfSZPM2iGq7ak8Etf+tJU2KwG/gi2RJzktiTu+c9/fjEOEFyMiWkTmrHeJUGDKJLo+pd/+Zdi/2iGa9999y0uKSz1a22Gi/Li0kHG2cte9rLkV3kZ6mCGElz8Ne1AVGCzWYW+ZnarRE28ILQGjROEGecs1jtdGrsWXNMD79zQJ5Ny6PkigKVm9H1ehgVXMxZcZprRLBVOZHSSeZ+EpVaM3TwOcBB5P2VaxsggwaVZJ8Z1qT0iLie8+OKLk8OM78Hz9c4771y4homrvcMFEi3vete7FsQqsym8x9N2p8OlgjYjPrBTdP4RHJtvvvnCmMiFgcbENAqGQYKLujN7RdtY1ld7R0v50K8aZ/Rr3DBFXHLJJdV3uOJyQs7D5jzHP/OZz7Re+rqcseCaIMz+YC+OWUSiqGmmqgQXd5t0CLmamBs3FlzTBcv8eEjQJ5O4SWv5HflbcFlwmflCMz/5DA5CbMcdd0xjtyTG3vSmNyUHc1oYl+DSeVpiyVJJftw2zuThYA/apRDiDx5TLk72oPeEpgEtKeQZH9uIKNhpp50WxkQuxhAq07rV/bgEF8QfPKZfsVfer4yRpl0KKW+DDTZYWLbJ8ktmuOI7XdMM7cMvZ8ZP47t28J41v002zufJ2AQXzrkEBrBcSLM7LDEb1tElH70npLzyZWyR0vmUzXtDWsLGZ+J0xLo1xQm9g6RzEA2cF+skW+VHtB0XE3Uq2RNqbS/tEjgJ+6sN5N0FhBZlUocmVN9xznLRB7SVPlHbqc8w73BFm3JTop7kx8GNR7Qdo8oPu5If/a96agzlUCbnKw0bLnB+/h4Peau/m/KjTogd1ZV8sA/1yh84uS05KCO2HdQubmDkgT00e0U5udOOAFB+8VAbAUdKGyFQfrTbICg/bqIQD9UlF1y0QQKN8nCESpA39tOmF7SPtfRd3qmCXHCVys/bynWvsksPAfor2o26/dVf/dVCviXxNEyauCOg6ls6L+7ESH1JN6iNbcjL33rrrVP5+hxtQ/uoB+XmdSSfaE9dG+rb2K/qd5XR1OcqL9bviiuuWKFsymLLbuyueukzaRgXt9xyy0r24V013q8jb/qKjSK4Nq677roVzuU8bITtS44Q7Xn729++0jjm2s7L5FzuOzhZ11xzTWr3Bz7wgfSZtLxr9JOf/KR/9uJQElzUnW/mee+kJMaIYzYnb19XcECx76tf/eqFPs4P7pux7BqDBBfPCOK75LlcKQkujQm2xieOay1eD/kOfOOA8YG4ec1rXrMgbvIjr0eNQYJL8V3yXI7QJ/S1rqU2Bzb/+c9/3s9hPIxFcEWHlgPHgM8MfOIU3tWx1qYSCBzgIcXDhbzIn88ROcSUqzgGJGGkifUmXu3h/AgPGByCUpzKj3WSo0tcBCeKcI4I+WMXlZHbE9R28lRb1D7SRFtOwv7UUelyOzehemP3QVA/8sd+44B60m7Kx/YgW6gteX/WyG3Kg7GUD/njtNTGKM6JwuI4pC9Ix/8K56Bc0BiJafI64WSB8tJn0sqxpdzoYMgB/t73vpc+y9EnLJ5HnQknvRxLHEI5ehJdOGKxToSTTmNPDglLDiKUq7jcAZIAiO2Rg93FWSIf0uRtgxhH3bALdeaz2iIbCfX1xz/+8ZQfNpITrr5ui8rHsc7LJ5zy+TZOYOcYlzsI6i/6RnajvhoHHLngIA3Oe5c0iAlswC6JhMexhsOusHwnRrWR/LjeS21sA+XjxEjw0AaNS9VZtomCivDYFupHfZQOccNnxCbjVmkkQlimVorL+wERgz3ULuogYaJNO1Qvlf3hD384XWccjAHSE8fzITpQpKPtyoe2M44IQwhF20fhljthcUMNCSXsmtcTsG0UtwgWZj+xBXWNomuxxADOE+2iTrxPQl0Jow58464NNbCN6sT9m6V0o9aRWTJmf9g98Kabblqw07A0CS7apGt+ww039A8ONyBb0e96140wrilmZiTQoyjBRxn3UkJmzJjxYXyUNmzpSpPgon3cs2gX4+OHP/yhx0cB7MQ9CztxH2dmjhlg2YoZXt5Va7Nl/qiMdYaLOA4eShE+19I1wQOHm3pOSQzhKORhAoe0VD7n1tKoznkcYRyRmrCqhYtavXiQ4gSV7MUFSBriebiKcdtfdeNoA3WRPXE+Y91qDLJPVyR8JQYEdZHDUOrrGrF+3LgVxv+MN8CxKeUphyfGYVM5UvQJ/Qz8xalRWbH+sin9jaMP5EMdcDjVNjnMIrY5OtOlcwnnxh1v1jiEjJf8Bi5HMcZhE5WFIxbT4Pzl50NMk5eBQ1sSB9g0P7eJNoILuyIGYrwEhISV4Gad1wsI5/xSXI1hyqdPeWgQlzuOcooRDpGYJq9f1zQKl9gS8fwY19TGYXZBZJxSDrM6efujgKnF5WUpP9IwbqNgjrseStwJCR/iopiRaMlneyQSscPDDz/cD/3tsk3uB3E2i7GuF+GjPdnim7Gct49wrunYNpWZCy7yRmyVhBhtplzqedddd/VDn0PfDJMutk9piFusd2Die1oSXAiX0047rffoo4+uJMbGtZQQx42849bro1ISXDjt9N+WW26Z2sHzdBaW9C0l8T0tCS58JcYE76Jx3XA9aGOJSSwllADivpdfW8NSElw//vGP07t4W221VWovM+i33Xabx0cFbSxCf5dmCRkLXGs8HyZ9D+vk6TJoubHyN6fJeVZcKV0NOfxySiM4m8TxoBKqW+n8WvlyZksOcy2OMrkB5nAuR6TJJlCrlwQSNiihGQWJABi3/dX+Nmli2Tp4aNfqL2I6/h8F5VWbLdP4KPV1jVi/0uyFxmgpLo5R3QjJTw/Q/ObIuNU351Ewqx9KaQBnshYn0UddiFf52Cg/H2GkMBxozsMZyKHNxJFvnCVTWJ6v4rB/jGtKQzh2y+0a69iGNoKrKS466DhAhJWWTsmhp85tHTGVgQPXpnwhRz+Wgy0JR4jk54NmcmJ+g9KU3uHSWCu1UaITcaLzYxvzNE1trKE65QIRogjqUpbEQslB6hqHDfIvHIDrSefH2SO1hxmuPI3iouBS2M0337zC+Yge+jOGYQ/KxPaxjiwFxE61LfK33377VEY+eyfBpZm0iOIWW3AhbplF4DOzW4itPI66jmspIfdDtlkv7Xw4LBJc2C8/Xve611lotUSCi35n23dmMBjDiFjGRIzDnpNYSog44norOfXDIsFVGx/f/e53PT4awH5rrLFG4/JAzYDx/P7BD34wUXtOreCSwzroAB5o/I8zUKJWfk1UQVOcwElmKRkdxbkckSabQK1eiBXCiS+hunGeaCqrVk4Tmi3qkoZvjPN+o29qcL7Oq7W1LSq31l+0oym+RLRp6SJsO0aVlvxwZqhLKT8cWs6P8errWhqJqqYjppWoIyyf6RJRVDUdSq92EZbXsdbmpjSa7eO6YmZkWORol2zXJi4XG6pv09H2Aa4yxiG4VLeaeCkJrkFp5NzHeImqpiOKq9jG3C5NbSwRZ9FKNl5qwRXr13REcdVVcLG8jzK475Om9EWPqAkuZreoR0k4AbM3KiOmmybBxQoAfvx4o402SrNWCCot78sFF+/OjWMpIcSdAMdFPsOF88c9ib73MsL2MCbYPIIxce+996bdFzUmcjGGzce9lBB4h2zcm0fkM1yaReMa9TLCZiSksNXnP//5xn7RLNig80alrAQqcCPgxsrfnOic5iiulK6GymrDoPxr8XJmS054UxwPV5xtHkrMRkjwcUSabAK1eikN8SW4CPN8m8qqldOE7N8ljYhCqmS/iM6rtbUtqm+tvEHxJaJNSzc1bny1uBLkx4OUupTSaMzF+FJYpCm/EvSNlqwpX8ZTRA5qTZDlqF0lW9Ta3JSGLzLiEkuEV768rg1qR8k+beKig66+HpezpTLGIbhKgipSii8JqkgpnrJL4qlGbGOepqmNJZoEFSy14FL5JfFUo6vgAtqidvL8YUatNOtaE1yqd01w1ZYITpPgwmGmjjicCCqWjanPoxhDcLHcctSlhIDzhi3pE9o66OC+3eY6yQUXIBD4PSzKatqZ0PwWnmHYkTHBzBVjQnaLYuyGG25Iy0vHLWSjECqNh/zgXtSmX3PBBYwP7bI3aGfC5YzsRL83zVxJmK2zzjpj35Uwp52i6YNzwmDhb050TnMUV0pXQ2XhIA4iOqUlauUrHX9zanF8+8+DDsEVly9yLkekySZQq5fS5I6wKOXbVFatnCaGmeGKDJpxAuynOlPHUdCDsFYe7RhUn5xo09LFKic8vnPVBPlRT+pSyk9jLsaXwiLkx3is3UxqMI7j7Bj9pTzkoOqdsUGoXeST16PW5qY0gnNkYw5msJu+1c9RO0q2axMXHfSufT0IlTEOwaWwmngpCa5BaWqCi7HW9oEU25inaWpjCZ2ft11Mi+AqvV9WYxjBBXzBF2dc6ZN8meEgwZXvmiiUjnOmXXBxPzjkkEPSRhkixh188MFjWUoIElyTnuESEhDYlWfAOMtcDHi2v+Md70j3Hr4UaCMuRkH20phAZKvfJVo0Jsa9lBAkuLjXjLOtJcEFCmd8cF+btH3HCTs4ssMnz95xbDxTA+3AFvZNywmBL5kGLTscF2UlUAED0cH8zcE5Io4jR3GldDXk8Dc5yIprKhtq5cuZLZVRitPMDRduDuEckWHrpTTxXZ6I0vHNv2gqq1ZOE2p/qa1t4IZA+pJtRawz/48CbSOfSb3DVbopdBmjQH44SdSllJ9sHsWPwmpptESwaTaKPGo3NZbEKg8JLMKa6gn0r74MUbvIIz+/1uamNDmcK3FYe+eohBztUjvaxEUHXRtMNAkE0nWt2zgElwRVl3e45NB3eYeLJTmUXRIBItogtjF3cJraWCIKqml8h4ud/ygf8VOacQKcUJYFqg7DCi6B8NJ5lBvrOUhw8bs6JadT6Xi2xHTTJLj0o8eUqY0xhAQXceOeHWJJIVvOjzPPmuAC7azHdcfMTG0cTCOMI40LjnHs2NeEfvSYsrRphojiJP8drnHCkkLGxzjzrwku0DI4nt96N20W2GOPPRau3ze84Q0Tm6GT4Gqa4dLsFvWZ9HJCKCuBCjgnGIm/OThFxHHkKK6UrgZOH2l4kJRmuXh44ZhCXNLHN/c5tfKjg5ujOP6KUpggnCPSZBOo1UuzQ1FQRWSbWI9x2x871vJrgwRXqe/EIPt0QTbjyLfzBtpOXKnvasT6lS7WOEZLMx9xjAL54RxRl1J+eocrvrekMVdLo6V3COPS7A9lUk/9X7o+CIv1os/4nNclQrmqj9rF+Xkda21uSoMd8jCu8dr5NeRol2zXJi466KUt8SP0daneNVTGOARXnO0obbNeElyD0pQEl0QnY622jI2xpvNjG0cVXPS/2h435hBLLbhA70flG04I6sHYUR26Ci62nC/ly/Izyo1iqCa4yIN8c0El9J4YdYplTaPgKgkqxbG0bBxLCSOII5YrjlM8NAkunEHuT/RXKX6a4X4Yfzz5q1/96thsVkKCqySoiMPGjAne75pUPfB5+PHpcYqfJsGlWTXGeyl+WrnkkksWBDCb0Dz++OP9mPHSRkx99KMfTXU54ogjZndb+BzF1QRECS5YnBvS8RenkDAgPx76/BWqG+fGcNLU6i1RkdeLQa40pR3j8vPleHOARAZ/FR7rLmSXfBYJ50Jtl7McoW7EK09QXhw5iutqf+UXy4lQz1ocdSwJ2YjsFvuF/LAH4bUZvhLRISd9rBf/03bixim4yFezQ23GKP+rjrk4Ul6kiWUNElwxT9LyABfUh5khlcW58bNQHvSXyuBGTp4cOKrYF/iL4029dG6sQ15HxWH/GBdFXawfEF4SNaV8mpBIijZVGXLCS3ZVXHTs875GiMa+5r24tu+8gcpoElwlYcHDgzpEwUOfKJy2xv6ljtrsIgqOPE3Mr5YmihqNNcWxiyNjK+YT21gTQaU21ogiMd9lT31NXK2sfAdB2qn2lB7GbQRXFKuUo/yYOZS4ohziuKZi3doIrnxXw9K7jDo3buUuwcWW0bH+9LuE4VlnnbVSXiz1YZzH7euhjeCifMUxHiifb5fz5Y6jwswPbSgJH80K5TNfOfQNW0F3EU84cIzlnXfeeWw7FTYJLqCeOPHYd9ZEF/z1X/91mqGblFMtmF3iHZyS2NEPImtL+CYkcLB3VxEjAcT4GJf4aRJcwPjg/TSu/1kSXSzdW3PNNXtnnnlm8f46LvAzmOViDHJPUP9z3+b30riPcB+cZB0irQUXToVEAIccFxFFRx6H46w4iZE2cG4sMx65M86DJJ6LQ4AzxcGgJSw69sCDUGmIw4nkL46qHF3iWTpGXWiX8tf5lCPhxkG94rI25U8YcaQRWpLGkdsl5inRRX1xiskzP38S9lf9sF8O9lae1InPCifdILEFqhdtFdHGHF3qG22Gjcifgz6S4CI81rcJjQGO2G8R6idHPD8oO0LbuDESR500E0ddcNipWxQa9LdmsDiG2VUwCjCVT9nKS2VQdpwZJDxurhEP0seHWnR28zrqd7iIi20D2S1eG+TL+dQnzq7JaY7tGQR9o7JxtimH+qnNKidvt8QmYyaKl6a+Lm0HXqNN+cTl5Uc75463tq0njjypDzalD+PYR+CovFHT5Ef+XpDaSF5RmNBGiQTiakvwckjH73epfJxfBCH9yphUOCKDsUNdSCPxmJel2ZxS3Wmn4nJhwpp/xeWzZppBIi4e2DG3geqFOMsFtJZvEicnIQo0jQsJq3yJIJtFqB654Cm1jfpQd8Z3fj7LJSUw85mvuMkGeSkO4abyacO4HBqJntq7VG2W/XEPoc7YuOtsFbsgMnPDNTLqTBdtabNskFk72Zhxzntpv/zlL/ux0wtjik0LGDOTdGixI/4Yz4hSOW2X+vF80u6F2tmw6xJE0uHIMz5GXQZKu6g7fc91WZs50wwe44N7H++o8RMJ08y5556bBHL8AeJJwX0Xn0/XEAf3Xu4h4/yJhza0ElyqZOnIHWQdEjdNcW3g4SORQVpuONFBj8jZ5zzOJx0XvepYKhcnSs4FfyUuopiK8Fl1IT8JApWLw0KZQudzRKedc/Mjd+rJm3xVHvUriYU8H45x2F92y4WDiP3CofqRrg3Yl/TRXrSNfDhym7WBsslXdZK9aDNHbezklJwmjtLNoe0YpW7kSz1oF23kfNLl/apzY9kcpC3VgfM1Bjn4P+8HPvMgII+YJ2VHp1/QL7Ge/OVz7K+mOg6qPzbiQUL7o+PKTAnnyZ5K12UGSUjwUXccrZpdKZ+2leJiufQRMy0SXvS1nPs2DFu+RFgMz2fH1L+Kl1AiLfYr1XOYNDj5ElOkwaGmjjovzoTFgzZyXi2ujQ0Ze/Sp7E+/UkfagdOKaNRYrtWDZXole3LgsMUHc5u4fAYPQYNNyZ8xHAUrxJmweFCvpjhsTp/oeuTg/3y2r1THXCghohBl0Y7UM4pCiMIpHjihTXE4M7QDB5BzujitTeCEnnDCCUWHkjh+jytuolFDwmkY0UQ5LEXaaaediu3n4FouCUIgPWOmljZ/BwkkzHROk0CbFpjd4tkyrr6vgT1PPPHEovNMHHWIm2jU4H4RhSxChiVvXWeNKPPv/u7v0vgoXcsc3O9qIpT03J9qaUszdXqfS+c0CbSl5p577uk8uzwvtJ7hmmV4GDMIeViZbugBP24Qk/RJkwDiRtlWIM0Ccrax6XK70Rhjlh84V7xjNskZjmEYRXCZwVx88cXpi5Fp6/cu8OX7gQceOHHBuJxgKSGzhpPcnXCaseAyjfCNMt/Sjlv4IKaYhalBuQi9OJsy61hwGWOWE8x0Nc32LBUWXGYQLOebxBbyZvliwWUGomWXWj45KnxzxFKsmpginGWM9Ns8YcFljFkusESL5bfT6LBacJkm4vtcxoyLZSW4JrE0brmg98lGFV30BfnUxBb5841ofJ9pXrDgMsYsB84555y0KmJa73MWXKZGl/e+jOnCshBczJYguDjGNUuzHEEkjSKGmNmap3eyuoDteOld4zDfsc8YY8ziYMFlahxzzDHL9h0jM1nmXnDJwY2HlxaaxUQzW6Vx6Ju6McYsLhZcpgQ7E7LtvpbB8iUxPxnh97jMOFgWM1zGGGOMMawuWHfddRe++Br0I8lmeaAfSI5fim644Ya9e++91+PDjAULLmOMMcYYY4yZEBZcxhhjjDHGGDMhLLiMMcYYY4wxZkJYcBljjDHGGGPMhLDgMsYYY4wxxpgJYcFljDHGGGOMMRPCgssYY4wxxhhjJoQFlzHGGGOMMcZMCAsuY4wxxhhjjJkQFlzGGGOMMcYYMyEsuIwxxhhjjDFmQlhwGWOMMcYYY8yEsOAyxhhjjDHGmAlhwWWMMcYYY4wxE8KCyxhjjDHGGGMmhAWXMcYYY4wxxkwICy5jjDHGGGOMmRAWXMYYY4wxxhgzISy4jDHGGGOMMWZCjCS4vv/97/cOP/zw3ote9KLe9ddf3w81o4BN3/e+9029TdXv3/72t/shs83Pfvaz3jbbbJMO/m/LZz/72ZTmVa96VT9keqGuz3ve89LfaQJ7b7fddsmOP/nJT/qh0wfX5tvf/vY07rHjxhtv3Lvooov6sbMD1+xaa63VO/TQQ3tPP/10P3R+eOyxx3rXXXdd74QTTujttttuvSOPPHIm28l1ccopp/Q23XTT3oc//OGxtwE78YzBTq9+9auTnZ566ql+7PIG259zzjm9zTffvPe3f/u3vSeffLIfM5h3vvOd6fq64YYb5vL6MsYMx9CC6xOf+ERykHA8OCy4Rgeb7rXXXjNhU9XxpJNO6ofMNti6i91xvukrOd+zILioo+r6zDPP9EOXHuy96qqrprpde+21/dDpQiJF9Tv77LMX6jxrogsnXnXv4kjOAg8++GBvs802S23Tcfnll8+c4/uNb3yjt/vuuy/007gFV8lOl156qQXCs2D717zmNQu2P/HEEztdJy94wQuGSmeMmW9GmuHiWyDdrGdFcPHt/rTUlVmiErNgU89wPQezkfTVNAmu2rha6hmu2rWHvad9hotZBvo4OqQSXVdccUU/ZLq4+OKLi/ae9xkuuOWWW3rrrbdeb6ONNurdd999/dDZY4cddkhjbBIzXHDrrbfOhZ0mwR577JFs31U4eYZrdvnnf/7n9KWa+81MgpHf4ZoFcRBhGdA01JUZEuxWYtZsupxhho++mhbBhaiZJvEXQbRM6wxWE1yHOF654Jp2ZtXe4+DKK69MfbbvvvvOtPOkmZZJCS7stPrqqyc7eTZmRU477bRkG89ULR/+6I/+qPflL3/ZgstMhGUluFiyNy111dLBErNk0+XONAkuZor4QmEaBRfXHo7jLAoALcGblNM7CXj/ZFbtPQ4OPPDA1P5ZXE4YmbTgeutb35ry93LClbHgWl588pOf7K2xxhq9f//3f/e1YCbCshFcWk41DXVlyZfqUmJa6mkGMy2CS0sip6EuOVx7OHXUzYJr8rCUcJbtPSp6P2kelslNUnDNk50mgQXX8oGlhHr3zoLLTIrWgivuSMig5Jt0zRhxlMQBjlbcWIP/WcIQwVEkH/LTBgz6TBqcx9p7Qnmd+Mvn/Hw5xaUjQhviphWxThHqHMulXbzLQ9ggaI/yj0d0khVGfShLuxYSRhmElWhj7ybatgu7YxfsUxOFXerCudhdNqD/1A/UJfYB5cmGeVxOqQ6MrRKUqbbXyPOjjqTh/9h/g8AOcZxRJvlg1xKla492axwQr+slP2QfzlX9o83kmJfSAPaO58Q4bBntkdcL+Kz4/NBDLdq+9KAjP/KJbcR+pfGELXS9IDZI+/73v3/Bdocddljr98TYtEDl5QcOcG67WPemOGz6tre9LZ1DOHXeZ599equsskqqZ5NzjZjifTfl+8Y3vnEFURU3xMgPOY1XXXVVKo8xW7O3dsdTWsrhXbX8fOr+gQ98YGFDEdLqM+l4T2yx38tjXGCDk08+uWrHLqhN0R6MezZLyfOnfxg32BZ7swGD+ovxW7Ih5DZnHLz3ve9NaWnLJAQXdkJQcG2NO++uXH311WlMauzSfsbOXXfd1T/jObCTdg/EJowtiVLOj8KIvtA7cOT5kpe8pNOMZxRclHPsscf21llnnZTXLrvs0rvppptWyuuee+5J/Uj9vvrVr64UzyxKrNOGG26Y8v/xj3/cP+O3YBOdiz24BrlPfu1rXxvYBux07rnnpnp85StfSfXn3TLaQ7mvf/3ri/UnHe0kXbTbBRdcsJLoRKRge8b6b37zm94HP/jBlD/nP/DAA+mcrvm9+c1vTvn9+te/7n3nO99JS13Jk/azm+YTTzyRzuW64h27UlyE8unHWD5tj+/XnXHGGWlmi7j8YCdPQZmqD3Hqu8cff7x/xnNEu9CO4447LuUvu7Bh1b/+67/2Xvayly30LXVq27dmdmkluHCIGBQ8ZHjAAjdrwjQwc8ebwRPPJ17n4/gpjPOUBzd+PnPwv5w5HlQ5qhPnclEB+eZlRFROXldQWsVRb5VPGREuJI7YNs4lrA2cr7qUUBw2Jk8cyFx05RA2yN6DaNMunGzKUh1lr4jiJW7on5hGBw9E2iVHmnKoK381FnQu4Rxy6vO4HOLz/pTIoU0aM0AdKFP5lcjbRH6xDtFGTVBXziet6kB7CKNt+e6BnE87KJfzOWI7Isonrws2kD2I5zxBO2Lb2WwghzJ5MEQ7Ixg4/+Mf/3j6zDl77713CqOsvB162OUzLoihWH7JAdCGGnfffXcKoz1yYilLabBRFLI4t+RN/9JmiQBEV5eHWtMMl5z7Ut2j+FEcjjttUR2xN22hftRT5+fOPHbAkSet+gg76Px8447VVlsthUd709dROGCbvM6D7B032sDxxd6xDtQxt/dib87BckIcY5wqnE45yQcddFDv0Ucf7Z/VDuyB08u1I2cIO+LAkeeFF164EBaFJg45fYg9GDfRTnfeeWc/9+colcHzLTrmkxBcLCeUnRAC66677oKdfvGLX/TPmjw8B3BiGScSHqeeemoK454ohxaHl59lkE1wdrGvnh8cOP6AuGDMytakVftw9tvYUoLrmGOOWbgmuGZUPkdciokNqZ8ccoROLEeC58wzz0xig35HYJDfIYccsoJgoL446Z/+9KfTVv2cq/rk+eZEoaY68pn6cyic8Xbbbbct5EUZO+64YxrDKuMHP/hBb6uttkppqAv1JiwXn/HLAcKYJULkKT+9F0XaP/zDPyzmJ1FEfvQRooVxH236T//0T+k5tOWWW6b+z+OiiKM9O+20UxpXzOYCfYRdafs3v/nNFez4u7/7uwt1z+2LiFp//fUX4qgzeVNf+o5xp3ZonO28887JLrEv/u3f/i3dT3/nd36n96lPfWphHNC31It4/zTD/DJQcDEYuKFx8H9EziMHg0jISaydn8fJ4aIMHjSCc0r5E04epMvBAVIZPAQjpbyA85rCYxz1K51LeKk+JUirfEsojvyinWLbIl3tXaJru/SQy89X27ixR9RnxFEvzlOdcJRVTwkagfOmuCgUgM/E5fVTeF43oF7EkW8k9nWO6pfXDWSHtn3PeZyf143ZjTycOpXaLRvn59fsISQQ8/zoBz0QSoJLYyA+hHR+DKMupXBQeC64gHbW0uFQYYN8loQ0cm5zsaG8qHNM1ySOmmgSXFDLM9o1xkU78UCOcTjtxOX2ljOf2wH7qK3x/JLgEji4pTKgrb1jOjk8u+666wrpcPzUzugITRItk9t2222TI6Tydbz85S/vPfLII/2zB4PjS/twAmMbNCZq4dgwF804Xlzn+TiS082MQwSby7bjFlzRTgcffHDRToslujRDlc8IadYhD3/ta1+7YOMbb7wxXWfYnfsafcH4RgzIyRYSdqTL40pI4LDkMs48MPMiJ5q8mMWISKDkwkhL1uL5ElZ5+H777Zeuz3z2hPC2syASNozRmIYy11xzzVQmMytR0BLO+RKuIDFQC8cGf//3f59sjzBBBD300EMpP9pMuti2mF8Mj2JI+QnNEhGXz2adfvrpKe6Vr3zlCvm9613vSvezGAaE06/5+TXBhZBCICGGSuGkiXGaMcvt8o53vCONu/333z/1bV6vAw44II11C675ZaDgkhNXcjaBOI7o+OGE5g4t8ADR+TjdgsFHWO4IguJi/nLCYx4ROdU4mBHC8ryA83KBIORQKy/S8plvLHNKbS6hPDhKKC6vJ5TSdbV3ia7tKvULNDn9mn3I+1lll9J0jePhSxg2KSHRyoF9IgqPkB83To4STe0twTgir7xs0pNPtKcE0ve+971+yHNQJ8Zrns+guig+tz+oLI3zCGH5+GGWhDrkD34e8ORTCy8JACilk0Ci/Dw/QITgwNLXMb6prFr9mhhWcEEpjj6upVEcfag4+piw0kyRxFMu3JoEV6kMYKlhrRyQGMTe0SmQKCiVpbjoPE0SvjDgG2bEIYJHM1qyE+HM6LSFLxtw3nP7dhViUBpHyoflYiUbSYyMW3BpO3jswSxdtBP1kZ3GWWYNbaN+xx139EOeoybEJLiYeSrZjNkP+iuP4zrCEWY8tpnlkuBiJiXPC9GlvPINR7SdfC64EGLcM3MRVRJiyuNb3/rWCnnwJQbX76C6g/IobYhC257//OencrX8jzYxi4Nd89k2CaeS4GKpXC4egPw22GCDlF9sc02I1cJBcblIglKcxNDnP//5lfoO8cP5tP2///u/F2xTE1wINGbvfvWrX/VDfgu7GmJjzXKBBBd2+b//+78UFvmTP/mTNK7+4z/+Y4VyWGbYtm/NbDJQcElwMBBKEBfjuakprOmITp+czZIjWHJEcTTzsIgcy9xJJqyUTm1sOqITq/IJq9WhCdIo3xKKK+WdpxvG3jW6tKvULyDbR3sJxeWiTvYopekap9mo0vmCeI78SwSFRxAaTfk1tbcN5E9alR3t+eIXvziFtf2R4kF1UTx/c+TUE6+lZEA44qrpIYAAxJbxOsrPV95dBBfvdBCGs1kCWyldnJlrKqtUziCWWnBpx8EuTvcwgkvv0NTKifW++eab+6HTJbj4tps65u/q8C4GjlxXwZXD9ZAvHYxt6yq4tESuZvNJCS7sxBjJnXHsxLjQUsNxltkWRB8zE7QbG9cEV0kI0T8SQk1HKW1Ok+ACvYOTx9cEV4R7pt5FUzujkIizPeRfesdrEKpHaZtzBIlER20bdC2Ri0sH28x81SjlNynBFUVV0xHFVU1wSVTlaePB8kGVLcEVwyLnnXdeqht9e/zxx6e+bfuMN7PNQMGlAcXDtkQez18+l5y6GnI4S2kUF8vPy8yJsxiRWjrCKKctfOupWTSlpcy2yEYcJRRXal+eTnl1sXeNLu0q9QtEAchDJaI0udBRG0p90DVOoqJ0vlAbc5sRxhEZlF+b8nKwC+m44TJ7QztITz7RnlpmuBiCC3bbbbeV4vm/dj59Tf1pB7OgfNaDKX+AK7yL4NJyOZzNEtixlG9TWbX6NbHUgmtQ+SWGEVzR3qVysHdJXE2L4JKoKu261xTXBpw67MOXCmeddVa6L9I28oxt6yq4BgmqSQguiaqanWpxk4SxhX3WXnvtNMPKWBo0w1USQvQTzm4bQTWIQYJL72R1EVyIDs3oMfPDO2Za3pc75yqfOO6xnN9FeDUJLtByuDweG2J7vmj72Mc+lpYQl4RVW8FVyq8krCSc8nBQXFvBpWWGuXhqoia4fu/3fq9Ybo1BggviRh307dFHH52WG1p4zTeNgosHMwOCIzqCkTxeaUpL02pwgydNybFTXCxfZdbEQKx3RGF5Wwhj0HeF2Yn4rX5pOVaJWv2E4vJ6Qp5OeXWx9yDatKvULyLOCuGEg2aeEDu5EFMbOD+naxxjSOXUUN3z8UYYR0TnlsoHlVeLz0HUYlvqJ9uAyon2lODKlxTWGFQXxeftFpRNmVwLehcHxyB/nwfoY2bgGBuxP3nAU0b+kFN4ySmHUjoJAMRc7aFZyreprFr9mpgWwVVb6ldiFMGVL5+LyAGM+ZbChOJGdX7boOWEpR871jtLOEFd3uECLaVEaClfnDzalgurroJL9llMwaXlhKUfO452Wqx3uFh6tvnmm6d7YtxQZBTBVVui2YVBgqsWXxNczNwxu8N1HIUTYoExUHLOJdA0Trg3t515HEZwfbC/02B8hyoKoa6Ci3evmvKLba6Fg+K6Cq78XbAmmgQXdv/f//3ffkgzbQQX0LcsV4zC6z//8z9b9a2ZTRoFFzAQOPJZCaF4OYo4lBo8uWMtCI/5ydksOYIlR1Rl1t6Z4lzicWwjShfzAupaCo/UnFSgLcqjZqeI6sdRQnGl+uTphrF3W5raVeqXCAJQ53DQF9iwVEfZg/NzusZJVHDU7KF6IRoiShdRftihhOJL9cuhPhKyed1Up2hPCa6mvotxg+qieP7W0A522IajJLYZczzIcZDyhwPhpK+Fl5xyKKWTACiVI5QuLoNsKqtWvyamRXAxBmsvVOPMxTE1iuBqsrccv7hdt8KWWnDVlhMCX84Rd9lll1XbVkLLOSf1DpdExWK+w1VbTgiauSvFTQLG7BZbbJHGSL6RxTCCS+9Wca3k+QnKpF8HbU4wSHBJ0Jx//vkr1K8kuFQv3utp8w5XDs652k3bms4VbQWXdirUMsau73Dl4SLmt9jvcKnsppkp3pniSxrZpia4/uAP/iCF55tmRBBZamNbwSXoW73XRd8yy2zmk4GCSw5iLl4EcRzRUWTQEFab8cHhw2kTcjZLjmDJESVfwmp1wgkt5UdYnhdoMwccjdwRBs6XY8v/uaMOhJFHzdmNkIfqUkJxeT2hlK6rvUt0bVepXwRCuCaGS8geJdt1jYtLGkvtAezFkfe10kXU/lp+2Ja4Uv1ymtpCGHHRnhqX1LU0y0TZcZZ3UF0Uz98aai/XFtdDadw05cMDnrj8waTwLoKLuig8Cioh4Ud7Y7qmskrlDKKt4CrVsVQefVyrh+Jim9ROzi/VgXGMWIrhwwgubSxBuvz3j4B6kC+7EcZ0ElVLLbjYDr5pOeEws1ty+vOd4XDyaBv5xrZ1FVx6hwsb5dvFwyQEF9vBb7LJJtXlhIs5u1WzIwwjuIDZO+zJTFIpHqEQHe0agwQXm3Nwb86FXUlwNeWFWKC+0Tk/4ogjir8rpZ0Hm94PE02CCyefcvktKZX7x3/8x+n8XHBEIdRFcCEiBuUX21wLB8W1FVwI3Be+8IXJrtp+PYddAWNeNcHFefRd/B2tCOUj4FXGIMH1F3/xF0V76V0xhKB3KpxPBgqu6HDmDj0PYMVFR1EOGQdOo5w2HAPicoeQz5xbcuAUF51dnGqJjNK3/6QpOdRKIyeVenEOdSecAyczOrGUi/OpvDg3fhbKoyZ6ItFuMV+huBgmFBeXo3W1d4mu7VK/RFuB2tZFcJEHaUr1VPlNcbnwpmyF19pTGmuEc+RpyIdwxo9sKySK2tpYZcT+I1xjk/8pn/h4PuOSz6B+JSyiLxpiuNKAxsmgvtGGALSp9FBXPtglxlO+ZuUIj7ZSntrCnbhoZx40xOfCUr//hJ3zuvDDtqTLHX3l1SS4SuKohhzlmj00K8iueIK2KV3eLvpE4Xl+isttq9knDpxv5cf52CjfGp93YTi3ZG+VwTjJy5e9mXHJ49Se3K44I6RpElxRTFAXxgNjPt86fVi0FI4tl/P8mNFBYHz3u9/thzy3/JCZlXwr9hw5/XE5J3aUUJJQ0FhvI7hiXlznstHWW2+9wlIzytm8/3tf4xJcslNpOWHJTsD9WXUc93bxOKvK+/bbb++HPheuzRUQXIx33TMHCS6JG9IyjvUDv7omuZbyWaYSTSKJGRTi8tktkNCJs4TKiy8EorNNPtotEOccoUAa8mAJa16u8o6/n1VD5zLDlOdDfSg31lGCix33ZB9spiWNElaqY1vBleenbdklrJTfl770pSScFB5pI7gQj3FXQM0aYVs2p+AeRDmITbbX55qK40AzWdrZkHox7pQ/cYgubCZ7sjkHwvv+++9fEGKDBBf14l22XFC/7nWvS/WVPYAlmYzXpdrAxoyXgYILNKPEgSOAw0UYDkgMj45cTBMPHrLR2eQClEOLU8VnwXlyRImLRCEo0UVayi05xiDnGEeDNsQ8Y375EUUFzgphuQNMXnnbmojtwm7UB5Q/R+4YS5hw5EKzrb1rdGkXtlX+0YYQ658f5M2YUVsFeRBfqmtsV96nElYcqjNQb8oinL/Kk3MogzxzJFY4cttSrvqLg/S0gbxVDgftiPXIoV7KR9cR9qAdupbIg/85F2Ib40E+uT34zAOOeNLF+pCfrjP+/vSnP03hJagX58UvOSLkqXLUn9iB8yW4VL4eEtpxkPM4P8Zhb+WHkxGhTRJr2F1CQ2miyIFog9xJ5fqJ5bR5QZmxI0GFzeNuiEKONPEsPVN/UleVp3ZTHy1TIzwXKXpfiLhYFvWQHfIjOvBCOw5SLnbI+6JUBkR7k6/srTrnAonzNZuW2zv+Dld8/0n2Ihyb5c7gMDA7wwwXvyul7cX5LaCjjjqqt/3226+w5TjjgDrTzkE7Fsa6Yg/6FlFKewhXn6tv3/SmN6VwrrEoMrn+JN6Ii8JKW7HLHhpDlKNrlnDK7/JFQQnsxAwXdpKwinbKxRb9SxzfuDeJtWHBLpqRoq2MIRxMbMBf7IVo4nrCZlwHCGXuM5yfbyUvEAnqt3hgR4Rdm2uffpEYQnQ9/PDDKbxJbOEs42yThnrLocdpVziCgnbSv2xPr3CEEdcmaSSWKFdjRXmUBFQJ5UHe8V0mLbeLQggkwjifOI1BxAHnYzvKZqxTJ0SLxm0+KwQSZG3yo258WUJ+hOeCUiKNvPIfLNZvdOX1YGzpx4mJiwfCKd/mPc5kYXf6QoKpadfDuPU8ZaodxJWWIUoIxh0K6VuWeHKtqU8YS1r2ye+llcSbmS1aCS7AwYk3fy4SBhefcYRKjiYOmBxS0nBedKjJg7j8IK+mOMHDAEeCvImjfnkZEc5XGziP+kfIm/xUFv/n7eIzbZKDrKOp3BLYhnrLlpDnqQNK4ZwfGWTvJtq2q9YvEdKUzokH50ApjjpQn65xsqPgs+zBQdqSiMjbrHMj2CCODdWDMhhTXB/5eCoRxyB/JeZJr3zzPuOctv3KQy3Pu2av6667LsXn0A6+6c4fFBHsqG3rqTPtgr333juFUY9oD+IlXKi/BF+cudFBfrFs8mFGQXbDBqVrs5QXB3k1xdWopeFoEkpRqODU0145yqWHv9rbFAeygwQRY6I2QxTtHesjRyAeuVCq2TtvM/aRwI4HzgdxpfYQhxCjDRrT11xzTT/H0dHGBORLGaeeeurCb0xFEA/8eO2gGS4o9S022nzzzZNtEF+MxZptsdsgu+Nw4YzKnhJX2JExMK6ZQIGdJHSa7PT1r399hXBmwRAM45zlwrHEltSFMad38BD52I3yEK+I35IdazNdtDH+QDE2JZ82YkvQL1wLXEvKJ86aRTSLlddPjnIcm7RJeSBcyBvxIXGlHz5WuaRhBgeBVmprCQkuxBbiQXVDUNTy0aYZnIdIoj6MdZYy0vYzzzwzjeeS+KBN+UyXNs1Qflx3MT8JQQmLPD+WTpbKYqarFoeQkWihLPqF8YUtsGEUsRHGoZZsqq4RxgJfNqg9iKC43JV4LWOMRz7ThbAr9e2nP/3plfqEWS+uz9LviZnZo7XgMqYtOJg4fIgCHJF4IFAkXIwxyxscoVFnbYahi+Ayv4UvcZhJtPM3/Uhw1TbNMMYsLvZ6zVhhZoNjEHy7bYxZvvAFDDM4S+EMWnANB8sRS7tAmunDgsuY6cKCy4wNHChmrrSUrQbT/BZcxixfmP1mmZOWOi42FlzdYdUC75h4dms2sOAyZrqw4DJjQ+8hIaZq7xghtli/rHd+jDHLC97h4l6xVGILLLi6wWYbLBVnNzYzG1hwGTNdWHCZsaGZK0QXB8KK5YW8t6V3t+JmDsYYsxRYcHXj6KOPTray4z4bsIHDmmuumZ7Db3jDG1bYjdAYszRYcJmxww52iC0JLw6EGKILUWaMMUuJBVd72JmQ33OT2OILs/jZTBea2YrPXw7PdBmztFhwGWOMWTbw22PaEp3jsssusyNagU0ycuedH++97777+mcYY4xpgwWXMcYYY4wxxkwICy5jjDHGGGOMmRAWXMYYY4wxxhgzISy4jDHGGGOMMWZCWHAZY4wxxhhjzISw4DLGGGOMMcaYCWHBZYwxxhhjjDETwoLLGGOMMcYYYyaEBZcxxhhjjDHGTAgLLmOMMcYYY4yZEBZcxhhjjDHGGDMhLLiMMcYYY4wxZkJYcBljjDHGGGPMhLDgMsYYY4wxxpgJYcFljDHGGGOMMRPCgssYY4wxxhhjJoQFlzHGGGOMMcZMhF7v/wFY6p35b9LkfgAAAABJRU5ErkJggg==\" width=\"860\" height=\"151\"\u003e\u003cbr\u003e\u003c/div\u003e\n\u003c/div\u003e\n\u003ch3\u003e3) Spatial-spectral features\u003c/h3\u003e\n\u003cp\u003eThe hybrid convolutional network employs multi-dimensional feature extraction by applying 3D convolution to concurrently extract spectral and spatial features from hyperspectral data, while utilizing 2D convolution to further extract spatial features. The network then progressively fuses features across layers, integrating high-dimensional spectral features into low-dimensional representations and introducing 2D convolution operations to aggregate and optimize spatial feature representations.\u003c/p\u003e\n\u003cp\u003eThis architecture demonstrates significant advantages in effectively integrating multi-dimensional features from hyperspectral data, thereby enhancing both the accuracy and robustness of classification tasks. By adopting a hierarchical feature extraction strategy, it optimizes computational resource allocation to address the inherent complexity and redundancy challenges in hyperspectral data processing. The detailed structure of the hybrid convolutional network model employed in this study is illustrated in Fig. 6.\u003c/p\u003e\n\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n \u003ch2\u003e3.2 TPMFN model advantages\u003c/h2\u003e\n \u003cp\u003eTo address the limitations of the aforementioned models, the TPMFN proposed in this paper exhibits the following advantages:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Multi-path architecture with adaptive receptive fields: By integrating convolution operations with varying receptive field sizes (kernel sizes), the TPMFN effectively captures multi-scale spatial-spectral features in hyperspectral images, simultaneously enhancing representation capabilities for both fine-grained details and global contextual patterns.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Dual-branch complementary learning: The architecture comprises two specialized branches dedicated to spectral feature extraction and spatial feature analysis, respectively. This design enables concurrent focus on localized spectral variations and macroscopic spatial distribution characteristics in hyperspectral imagery.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Hierarchical feature fusion: Through a learnable fusion strategy, the TPMFN achieves feature map alignment and cross-modality integration after multi-path extraction. This mechanism produces comprehensive characterizations of hyperspectral data complexity, ultimately leading to superior classification performance.\u003c/p\u003e\n \u003c/span\u003e\n\u003c/div\u003e"},{"header":"4. Results and Discussion","content":"\u003cp\u003eIn this section, we first describe the hyperspectral image dataset used in our experiments. Subsequently, ablation experiments are conducted to systematically analyze the impacts of three distinct spectral-spatial feature extraction modules on model performance. Finally, the proposed model is benchmarked against mainstream approaches, with both quantitative metrics and visual comparisons being utilized to evaluate fungal lesion classification capabilities across different models.\u003c/p\u003e\n\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\n \u003ch2\u003e4.1 Data-set description\u003c/h2\u003e\n \u003cp\u003eThe hyperspectral image dataset used in this study contains mold spots induced by the six aforementioned fungal infections, acquired at a sampling height of 40 cm. Each image has dimensions of 521 \u0026times; 364 pixels, containing 19,951 valid pixels per spectral band across 300 spectral channels. The dataset was partitioned into training(10%) and test(90%) subsets. Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e details the class nomenclature and corresponding sample distribution between the training and test sets for the classification task.\u003c/p\u003e\n \u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eClass-specific mold coverage characteristics and training-test sample allocation\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNo.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eClass\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTraining\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTest\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePaecilomyces lilacinus\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e250\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e2251\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eAspergillus niger\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e358\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e3225\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003ePenicillium citrus\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e365\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e3289\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eAlternaria alternata\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e652\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e5867\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eTrichoderma longiformis\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e184\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1650\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cem\u003eCladosporium cladosporioides\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e186\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1674\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e1995\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e17956\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\n \u003ch2\u003e4.2 Ablation study\u003c/h2\u003e\n \u003cp\u003eIn order to fully validate the effectiveness of the proposed method, ablation experiments with different component combinations were conducted on the dataset mentioned in this paper. Six configurations were considered, and the impact of each component on the overall model accuracy was analyzed through classification performance. All experimental results are listed in Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e. Specifically, the model was divided into four modules: 2D-CNN\u0026thinsp;+\u0026thinsp;Spectral_Att, SpectralFormer, Spectral_Att, and Hybrid Network. Case 1 (excluding both SpectralFormer and Spectral_Att) achieved the lowest accuracy of 90.31%. Case 4 (without 2D-CNN\u0026thinsp;+\u0026thinsp;Spectral_Att) showed a slight improvement to 91.01%. Case 5 (removing Hybrid Network) reached 93.45% accuracy by relying solely on independent processing of spatial and spectral features. Comparing Case 3 (SpectralFormer alone) and Case 6 (SpectralFormer\u0026thinsp;+\u0026thinsp;Spectral_Att), the accuracy significantly increased to 97.12%. This demonstrates that Spectral_Att positively enhances spectral feature processing, thereby improving classification accuracy.\u003c/p\u003e\n \u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003ePresents the analysis of the proposed model conducted on this dataset (Suboptimal results)\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\"\u003e\n \u003cp\u003eNo.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"4\"\u003e\n \u003cp\u003eComponent\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003eIndicators\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e2D-CNN\u003c/p\u003e\n \u003cp\u003eSpatial_Att\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSpectralFormer\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSpectral_Att\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eHybrid Network\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eOA(%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAA(%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u0026times;100\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026times;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026times;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e90.31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e85.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e87.68\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026times;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e93.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e92.83\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026times;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e92.91\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e90.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e92.27\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026times;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e88.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e89.58\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026times;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e93.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e90.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.46\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026radic;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e93.24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e95.16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003eTo further validate the effectiveness of the proposed algorithm, thorough comparative experiments employing different feature fusion approaches were conducted on this data-set. Three distinct fusion methods - additive fusion, multiplicative fusion, and concatenation fusion - were implemented, with all experimental results summarized in Table \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. As shown in Fig. \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e, it can be observed that models employing multiplicative fusion generally demonstrate superior classification performance compared to those using additive fusion and concatenation fusion, achieving a top accuracy of 98.84%.This indicates that the multiplicative fusion approach exhibits enhanced capability in feature interaction and information coupling, enabling more effective exploitation of complementary relationships among multimodal features to improve classification performance.\u003c/p\u003e\n \u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003ePerformance analysis of the proposed model on the benchmark dataset (Bold entries denote optimal results)\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNo.\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFusion Methods\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eOA(%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAA(%)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e\u0026times;100\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAdditive Fusion\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e94.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e93.26\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMultiplicative Fusion\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e98.84\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e98.52\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e98.54\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eConcatenation Fusion\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e96.67\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e96.81\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\n \u003ch2\u003e4.3 Quantitative analysis\u003c/h2\u003e\n \u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e presents the OA, AA, Kappa coefficient, and per-class classification accuracies obtained by all methods described in Section \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e. The optimal results are highlighted in bold.The evaluation data clearly demonstrate that the proposed TPMFN method achieves the best performance, yielding the highest OA, AA, and Kappa coefficient values, along with superior classification accuracies for specific categories. For instance, in the \u003cem\u003eTrichoderma longibrachiatum\u003c/em\u003e class, models including SVM, 1D-CNN, 2D-CNN, Spectralformer, and SSFTT exhibit limited effectiveness, potentially due to the small sample size and non-concentrated distribution of this class, which hinders feature learning. Furthermore, the percentage-based random sampling strategy may exacerbate class imbalance issues.\u003c/p\u003e\n \u003cp\u003eIn contrast, TPMFN delivers consistently high classification accuracies (above 96%) across all classes, indicating its strong capability in handling class imbalance for hyperspectral image classification tasks. Whether for dominant or minor classes, the model maintains robust performance. This consistency likely stems from its multimodal feature fusion mechanism and efficient hierarchical feature extraction.\u003c/p\u003e\n \u003cp\u003eHowever, there is another case: in the category of \u003cem\u003ePaecilomyces lilacinus\u003c/em\u003e, the classification accuracy of SSFTT is significantly higher than that of TPMFN. The main reason is the highly convergent sample distribution of this category, which forms a compact, nearly circular pattern, whereas the sample distributions of other categories are more dispersed.Therefore, the proposed method does not show significant advantages in classifying this category. However, for datasets with discrete and localized sample points, TPMFN demonstrates a significant advantage, enabling better capture of fine-grained local information.This paper also investigates the impact of different training sample proportions on different models, as shown in Fig. \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e. With a small number of samples, TPMFN still maintains good performance. As the sample size increases, the performance of SSFTT and HybridSN is only slightly lower than that of the proposed method.\u003c/p\u003e\n \u003ctable id=\"Tab5\" border=\"1\" class=\"fr-table-selection-hover\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eClassification accuracy of different classification methods on the dataset (Bold data indicates the best results in each category)\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e序号\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e1-D-CNN\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003e2-D-CNN\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSSFTT\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSpectralformer\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eHybirdSN\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTPMFN\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e34.18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e46.38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e85.66\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e98.58\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e88.31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e83.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.46\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e44.65\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e59.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e87.13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e98.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e85.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e99.06\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e60.18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e83.37\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e92.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e99.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e96.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e100.00\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e74.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e92.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e93.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e98.46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e98.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e99.45\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e27.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e34.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e61.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e58.64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e63.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e85.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e96.54\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e25.17\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e62.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e48.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e96.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e99.28\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e99.16\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOA(%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e61.18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e70.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e85.09\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e94.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e95.38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e98.84\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAA(%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e54.47\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e61.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e78.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e90.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e88.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e93.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e98.52\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026times;100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e56.09\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e62.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e80.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e92.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e89.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e94.17\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e98.54\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003e\u003cstrong\u003e4.4 Visual evaluation\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eThe classification maps of the aforementioned methods on this dataset are presented in Fig. 9. As observed, the TPMFN classification map exhibits the cleanest spatial patterns and closest resemblance to the ground truth. In contrast, conventional methods including SVM, 1D-CNN and 2D-CNN demonstrate limited capability in capturing discriminative spectral signatures or deep spatial features, resulting in noisy outputs with substantial misclassifications. This fundamental limitation ultimately leads to their poor performance in target recognition tasks.While advanced approaches like SpectralFormer, SSFTT and HybridSN achieve classification accuracies exceeding 90%, persistent errors remain in complex scenarios, particularly within regions containing \u003cem\u003ePaecilomyces lilacinus\u003c/em\u003e and \u003cem\u003eTrichoderma longibrachiatum\u003c/em\u003e species. Remarkably, the proposed TPMFN method successfully identifies these challenging mold colonies with significantly higher accuracy, quantitatively validating its superior performance.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eIn this paper, we propose a multimodal feature fusion method, TPMFN, based on hyperspectral imaging to address the accuracy and efficiency challenges in mold detection in paper artifacts. TPMFN integrates spectral, spatial, and joint spatial-spectral features within a multipath structure to achieve deep feature extraction and fusion.Specifically, TPMFN captures spectral details using a Spectral Transformer, extracts spatial patterns of mold spots via a combination of 2-D CNN and a spatial attention mechanism, and further integrates multi-dimensional features through a hybrid convolutional network to enhance feature representation capability and improve computational efficiency.To further explore this capability, we investigate several fusion modules, including additive, multiplicative, and concatenation-based fusion. Experimental results demonstrate that the multiplicative fusion method significantly outperforms traditional classification methods and single-path networks in terms of accurately capturing local feature variations of mold spots, enhancing key information regions, and improving classification performance.The innovative design of TPMFN offers a novel perspective for hyperspectral image processing tasks and provides an efficient and intelligent solution for mold detection in paper artifacts.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eThis work was financially supported by the Chongqing Talents Program [grant numbers cstc2021ycjh-bgzxm0287].\u003c/p\u003e\n\u003ch2\u003eConflicts of Interest\u003c/h2\u003e\n\u003cp\u003eThe authors declare no conflict of interest.\u003c/p\u003e\n\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\n\u003cp\u003eDeng Xuexu, Zhao Ya and Qin Dan contributed to the study conception and design.Material preparation , data collection and analysis were preformed by Ma zheng, Xiao Zhongyu, Luo Xiling, Song Tao, Tang Bin and Wang Jianxu. Zhao Mingfu and Tang Huan are responsible for the supervision of research activities.The first draft of the manuscript was written by Deng Xuexu.All authors read and approved the final manuscript.\u003c/p\u003e\n\u003ch2\u003eAcknowledgement\u003c/h2\u003e\n\u003cp\u003eSpecial thanks to the key scientific research base of the State Administration of Cultural Heritage for research on pest control of cultural relics in collections (China Three Gorges Museum, Chongqing) for providing samples for this paper.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eZhang, Xu, et al. \u0026quot;Chemistry directs the conservation of paper cultural relics.\u0026quot;\u0026nbsp;Polymer Degradation and Stability\u0026nbsp;207 (2023): 110228.\u003c/li\u003e\n \u003cli\u003eCarter, Henry A. \u0026quot;The chemistry of paper preservation: part 2. The yellowing of paper and conservation bleaching.\u0026quot;\u0026nbsp;Journal of Chemical Education\u0026nbsp;73.11 (1996): 1068.\u003c/li\u003e\n \u003cli\u003ePinheiro, Ana Catarina, S\u0026iacute;lvia Oliveira Sequeira, and Maria Filomena Macedo. \u0026quot;Fungi in archives, libraries, and museums: a review on paper conservation and human health.\u0026quot;\u0026nbsp;Critical reviews in microbiology\u0026nbsp;45.5-6 (2019): 686-700.\u003c/li\u003e\n \u003cli\u003eYang Liguang. Causes of Damage to Paper Artifacts and Measures for Their Protection and Restoration [J]. Collection, 2023, (12): 110-112(in Chinese).\u003c/li\u003e\n \u003cli\u003eSchmitz, Kevin, et al. \u0026quot;Preserving cultural heritage: Analyzing the antifungal potential of ionic liquids tested in paper restoration.\u0026quot;\u0026nbsp;PloS one\u0026nbsp;14.9 (2019): e0219650.\u003c/li\u003e\n \u003cli\u003eWu Fasi, Li Jie, Li Ruixi, et al. Research Progress on the Application of Fungicides and Antibacterial Nanomaterials in Cultural Heritage Conservation [J]. Sciences of Conservation and Archaeology, 2023, 35(05): 115-127(in Chinese).\u003c/li\u003e\n \u003cli\u003eRivas T, Pozo-Antonio J S, de Silanes M E L, et al. Laser versus scalpel cleaning of crustose lichens on granite[J]. Applied Surface Science, 2018, 440: 467-476.\u003c/li\u003e\n \u003cli\u003ePfendler S, Einhorn O, Karimi B, et al. UV-C as an efficient means to combat biofilm formation in show caves: evidence from the La Glaci\u0026egrave;re Cave (France) and laboratory experiments[J]. Environmental Science and Pollution Research, 2017, 24: 24611-24623.\u003c/li\u003e\n \u003cli\u003eMeng, Qingxia, et al. \u0026quot;A biological cleaning agent for removing mold stains from paper artifacts.\u0026quot;\u0026nbsp;Heritage Science\u0026nbsp;11.1 (2023).\u003c/li\u003e\n \u003cli\u003eYan Li, Hong Wei, L\u0026uuml; Xiaofang. Isolation and Molecular Identification of Molds on Archived Paintings and Silk Artifacts [J]. Cultural Relics Identification and Appreciation, 2021, (10): 60-65.\u003c/li\u003e\n \u003cli\u003eStrycker B D, Han Z, Duan Z, et al. Identification of toxic mold species through Raman spectroscopy of fungal conidia[J]. PloS one, 2020, 15(11): e0242361.\u003c/li\u003e\n \u003cli\u003eVashpanov Y, Heo G, Kim Y, et al. Detecting green mold pathogens on lemons using hyperspectral images[J]. Applied Sciences, 2020, 10(4): 1209.\u003c/li\u003e\n \u003cli\u003eChun S W, Song D J, Lee K H, et al. Deep learning algorithm development for early detection of Botrytis cinerea infected strawberry fruit using hyperspectral fluorescence imaging[J]. Postharvest Biology and Technology, 2024, 214: 112918.\u003c/li\u003e\n \u003cli\u003eZou Z, Zhen J, Wang Q, et al. Research on nondestructive detection of sweet-waxy corn seed varieties and mildew based on stacked ensemble learning and hyperspectral feature fusion technology[J]. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 2024, 322: 124816.\u003c/li\u003e\n \u003cli\u003eSiripatrawan U, Makino Y. Assessment of food safety risk using machine learning-assisted hyperspectral imaging: Classification of fungal contamination levels in rice grain[J]. Microbial Risk Analysis, 2024, 27: 100295.\u003c/li\u003e\n \u003cli\u003eLu M, Zhang Q, Li Y, et al. Research on Labeling Method for Fungal Diseases in Murals Based on Hyperspectral Images[C]//2024 3rd International Conference on Image Processing and Media Computing (ICIPMC). IEEE, 2024: 93-100.\u003c/li\u003e\n \u003cli\u003eDai R, Tang B, Zhao M, et al. Study on Detection Method of Foxing on Paper Artifacts Based on Hyperspectral Imaging Technology[C]//Journal of Physics: Conference Series. IOP Publishing, 2021, 2010(1): 012177.\u003c/li\u003e\n \u003cli\u003eHu W, Huang Y, Wei L, et al. Deep convolutional neural networks for hyperspectral image classification[J]. Journal of Sensors, 2015, 2015(1): 258619.\u003c/li\u003e\n \u003cli\u003eZhao W, Du S. Spectral\u0026ndash;spatial feature extraction for hyperspectral image classification: A dimension reduction and deep learning approach[J]. IEEE Transactions on Geoscience and Remote Sensing, 2016, 54(8): 4544-4554.\u003c/li\u003e\n \u003cli\u003eSun L, Zhao G, Zheng Y, et al. Spectral\u0026ndash;spatial feature tokenization transformer for hyperspectral image classification[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-14.\u003c/li\u003e\n \u003cli\u003eHong D, Han Z, Yao J, et al. SpectralFormer: Rethinking hyperspectral image classification with transformers[J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 60: 1-15.\u003c/li\u003e\n \u003cli\u003eRoy S K, Krishna G, Dubey S R, et al. HybridSN: Exploring 3-D\u0026ndash;2-D CNN feature hierarchy for hyperspectral image classification[J]. IEEE Geoscience and Remote Sensing Letters, 2019, 17(2): 277-281.\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":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-heritage-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hsci","sideBox":"Learn more about [Heritage Science](http://heritagesciencejournal.springeropen.com)","snPcode":"40494","submissionUrl":"https://submission.nature.com/new-submission/40494/3","title":"npj Heritage Science","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Hyperspectral imaging, Mold spot detection, Multimodality, Feature fusion, Deep learning","lastPublishedDoi":"10.21203/rs.3.rs-6292446/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6292446/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAs an important carrier of Chinese civilization, the prevention and control of mold growth in paper cultural relics is the core challenge in the field of cultural relic protection. To address the issues of low sensitivity and poor timeliness of traditional detection methods, this study proposes a TriplePath Multimodal Feature Fusion Network (TPMFN) based on hyperspectral imaging. Our research constructs an improved two-dimensional convolutional network (2D-CNN), extracts RGB bands using spatial attention mechanism, and introduces improved SpectralFormer module to accurately capture mold specific responses in the 400-1000nm spectral range, ultimately achieving efficient integration and classification decision-making of multi-source information. In this study six categories of typical fungal colonization samples from paper-based cultural relics were analyzed. The research results indicate that the proposed TPMFN has significant advantages over baseline models including including Support Vector Machine (SVM), 1D-CNN, 2D-CNN, SSFTT, HybridSN and SpectralFormer, achieving 98.84% overall accuracy and 98.54% kappa coefficient. Notably, it attained a 3.6% improvement in detailed feature identification accuracy compared to the suboptimal baseline. The ablation study with component-wise evaluation revealed that the dual-attention mechanism enhanced feature discriminative power with a relative increase of 7.5%. This study provides a high-precision and practical solution for mold detection of paper cultural relics, which has important practical value for establishing a preventive protection system\u003c/p\u003e","manuscriptTitle":"Mold Spot Detection for Paper Artifacts Based on Multimodal Feature Fusion","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-21 10:07:16","doi":"10.21203/rs.3.rs-6292446/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-05-13T22:00:39+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-13T10:13:50+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-13T03:06:48+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"333404689410487198128915641209193370437","date":"2025-04-24T01:49:12+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"311099837615993966309663764440184710031","date":"2025-04-23T10:48:39+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-23T08:57:41+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"145739882112403600472515101087093145837","date":"2025-04-08T01:13:58+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-03-30T11:53:28+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-03-26T11:56:13+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-03-26T11:53:24+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj heritage science","date":"2025-03-24T06:54:56+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-heritage-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hsci","sideBox":"Learn more about [Heritage Science](http://heritagesciencejournal.springeropen.com)","snPcode":"40494","submissionUrl":"https://submission.nature.com/new-submission/40494/3","title":"npj Heritage Science","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"473b2093-7c63-4451-a620-c7f2209ff48b","owner":[],"postedDate":"April 21st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-10-27T16:27:11+00:00","versionOfRecord":{"articleIdentity":"rs-6292446","link":"https://doi.org/10.1038/s40494-025-02102-1","journal":{"identity":"npj-heritage-science","isVorOnly":false,"title":"npj Heritage Science"},"publishedOn":"2025-10-25 16:17:05","publishedOnDateReadable":"October 25th, 2025"},"versionCreatedAt":"2025-04-21 10:07:16","video":"","vorDoi":"10.1038/s40494-025-02102-1","vorDoiUrl":"https://doi.org/10.1038/s40494-025-02102-1","workflowStages":[]},"version":"v1","identity":"rs-6292446","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6292446","identity":"rs-6292446","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}