Method and apparatus for automatic pilling rating of fabric

Abstract

The application provides a fabric pilling automatic grading method and device, and relates to the technical field of textile detection. The method comprises the following steps: collecting a thermal infrared polarization image sequence of a fabric sample by using a thermal infrared polarization camera; calculating a polarization degree and a thermal decay time constant based on the sequence data, and extracting a micro-thermal radiation feature; simultaneously collecting a visible light image by using a visible light camera, and extracting an image feature; fusing the micro-thermal radiation feature and the image feature to generate multi-modal feature data; and finally identifying and determining a pilling grade by using a neural network model. By combining thermal physical characteristics and visual features, the problems of strong subjectivity of artificial grading and large limitation of single optical detection are overcome, and objective, accurate and efficient automatic grading is realized.

Description

Technical Field

[0001] This invention relates to the field of textile testing technology, and in particular to an automatic rating method and apparatus for fabric pilling and fuzzing. Background Technology

[0002] Pilling and fuzzing are important indicators affecting the appearance and quality of fabrics, and currently, the industry mainly relies on manual visual assessment for rating. This method suffers from strong subjectivity, and the rating results are greatly affected by the assessor's experience, ambient light, and fatigue, resulting in poor consistency and repeatability. Furthermore, existing automatic rating methods are mostly based on visible light image analysis, which can only capture two-dimensional visual features of the fabric surface. They struggle to effectively identify the microstructure, density, and distribution of pilling and fuzzing, such as the depth of the pilling and other intrinsic properties, thus limiting the accuracy and robustness of the rating.

[0003] With the diversification of fabric types and processing techniques, there is an urgent need for an automatic rating method that can integrate multi-dimensional information, be objective and efficient, in order to improve rating accuracy and automation level. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, the first aspect of this invention proposes an automatic rating method for fabric pilling and fuzzing, comprising:

[0005] S1: Based on the fabric sample, use a thermal infrared polarization camera to acquire the thermal infrared polarization image sequence of the fabric sample and generate thermal infrared polarization image sequence data;

[0006] S2: Based on thermal infrared polarization image sequence data, micro-thermal radiation characteristics of fabric samples are extracted by calculating polarization degree and thermal decay time constant, and micro-thermal radiation characteristic data are generated.

[0007] S3: Based on the fabric sample, use a visible light camera to acquire visible light images of the fabric sample and generate visible light image data;

[0008] S4: Based on visible light image data, image features of fabric samples are extracted through grayscale processing, texture feature calculation, and color feature calculation to generate image feature data;

[0009] S5: Based on micro-thermal radiation feature data and image feature data, multimodal feature data is generated by feature splicing and normalization.

[0010] S6: Based on multimodal feature data, the pilling and fuzzing level of the fabric sample is identified and determined by a neural network model; wherein, S6 includes:

[0011] S61: Based on multimodal feature data, construct feature maps and generate graph structure data;

[0012] S62: Based on graph structure data, initial rating data is generated through graph neural network model identification;

[0013] S63: Based on the initial rating data, weighted rating data is generated through an attention mechanism.

[0014] S64: Generate pilling and fuzzing level data based on weighted rating data.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0016] An automatic rating method for fabric pilling and fuzzing effectively addresses the subjectivity and inaccuracy issues mentioned in the background technology through a multi-step synergistic process. First, step S1 uses a thermal infrared polarization camera to acquire a sequence of thermal infrared polarization images of the fabric sample, generating thermal infrared polarization image sequence data. This step captures the dynamic response of the fabric under thermal excitation, providing fundamental data for subsequent analysis. Step S2, based on the thermal infrared polarization image sequence data, extracts the micro-thermal radiation characteristics of the fabric sample by calculating the degree of polarization and the thermal decay time constant, generating micro-thermal radiation characteristic data. The degree of polarization reflects the polarization state of thermal radiation on the fabric surface, while the thermal decay time constant characterizes the thermal conductivity. These micro-thermal radiation characteristics can sensitively detect microscopic changes in pilling and fuzzing, such as differences in fiber arrangement and thermophysical properties, overcoming the limitations of purely visual features. Step S3 uses a visible light camera to acquire visible light images of the fabric sample, generating visible light image data, ensuring the acquisition of traditional visual information. Step S4, based on visible light image data, extracts image features from the fabric sample through grayscale processing, texture feature calculation, and color feature calculation, generating image feature data. These image features cover visual attributes such as fabric surface texture roughness and color variation, complementing the micro-thermal radiation features. Step S5, based on the micro-thermal radiation feature data and image feature data, fuses them through feature splicing and normalization to generate multimodal feature data. This fusion process integrates thermophysical and visual information, eliminates feature scale differences, and enhances the representativeness and robustness of the data. Finally, Step S6, based on the multimodal feature data, identifies and determines the pilling level data of the fabric sample using a neural network model. The neural network model can automatically learn complex patterns in the multimodal features, achieving efficient and accurate classification.

[0017] The entire method is tightly integrated across its various steps: thermal infrared polarization image sequence data provides intrinsic thermophysical characteristics, visible light image data provides extrinsic visual characteristics, multimodal feature data fusion ensures comprehensive information, and the neural network model enables end-to-end intelligent rating. This synergy significantly improves the objectivity, accuracy, and efficiency of the rating, reduces human interference, and is applicable to diverse fabric types. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 The diagram shown is a flowchart illustrating an automatic grading method for fabric pilling and fuzzing according to an embodiment of the present invention.

[0020] Figure 2 The diagram shown is a structural schematic of an automatic fabric pilling and fuzzing rating device according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0022] The specific embodiments of the present invention will be described below.

[0023] Example 1

[0024] like Figure 1 As shown, this invention proposes an automatic rating method for fabric pilling and fuzzing, comprising:

[0025] S1: Based on the fabric sample, use a thermal infrared polarization camera to acquire the thermal infrared polarization image sequence of the fabric sample and generate thermal infrared polarization image sequence data;

[0026] S2: Based on thermal infrared polarization image sequence data, micro-thermal radiation characteristics of fabric samples are extracted by calculating polarization degree and thermal decay time constant, and micro-thermal radiation characteristic data are generated.

[0027] S3: Based on the fabric sample, use a visible light camera to acquire visible light images of the fabric sample and generate visible light image data;

[0028] S4: Based on visible light image data, image features of fabric samples are extracted through grayscale processing, texture feature calculation, and color feature calculation to generate image feature data;

[0029] S5: Based on micro-thermal radiation feature data and image feature data, multimodal feature data is generated by feature splicing and normalization.

[0030] S6: Based on multimodal feature data, the pilling and fuzzing level data of the fabric sample is identified and determined by a neural network model.

[0031] The automatic assessment method for fabric pilling and fuzzing first involves acquiring a sequence of thermal infrared polarization images of fabric samples using a thermal infrared polarization camera. This camera is an imaging device capable of simultaneously detecting the thermal radiation intensity and polarization state of an object's surface. Its working principle involves receiving the long-wave infrared radiation emitted by the object itself and imaging it at different polarization angles using a built-in polarization filter assembly. The acquired thermal infrared polarization image sequence data is a set of time-series, multi-angle polarized thermal images that record the changes in thermal radiation intensity at various points on the fabric surface during the cooling process after thermal excitation, along different polarization directions. This step provides rich raw thermophysical data for subsequent analysis.

[0032] Based on the obtained thermal infrared polarization image sequence data, the system extracts micro-thermal radiation features by calculating the degree of polarization and the thermal decay time constant. The degree of polarization refers to the proportion of polarized components in thermal radiation, with a value between 0 and 1, reflecting the directional vibration characteristics of thermal radiation. For fabric surfaces, the fiber arrangement disorder is higher in pilling and fuzzy areas, leading to a depolarization effect in thermal radiation, thus causing the polarization characteristics of these areas to differ from those of smooth areas. The thermal decay time constant is a physical quantity describing the time required for the surface temperature of an object to recover from a heated state to an equilibrium state, reflecting the thermal inertia characteristics of the material. Because the fibers in pilling and fuzzy areas are fluffy and porous, their thermal insulation performance is better, and the thermal decay rate is slower than that in smooth areas. By calculating these two parameters, the system can obtain micro-thermal radiation feature data that includes both surface structure information and material thermal property information.

[0033] Simultaneously, the system also uses a visible light camera to acquire visible light images of the fabric samples. Operating in the 380-780 nm visible spectrum, the camera captures fabric surface morphology features recognizable to the human eye. The generated visible light image data records the fabric's appearance under normal lighting conditions, including visual information such as the distribution, size, and density of pilling and fuzzing. This step preserves the visual features relied upon by traditional manual grading, providing the system with intuitive surface morphology data.

[0034] Next, the system processes the visible light image data to extract image features. Grayscale conversion is the process of converting a color image to a grayscale image. A weighted average method is used to fuse the information from the three RGB channels into single-channel brightness information, simplifying subsequent calculations and highlighting texture features. Texture feature calculation quantifies surface roughness, contrast, uniformity, and other characteristics by analyzing the spatial distribution of image pixel grayscale. Commonly used methods include algorithms such as gray-level co-occurrence matrix and local binary mode. Color feature calculation extracts the statistical distribution characteristics of the image in the color space, such as color histograms and color moments, to characterize color changes or shadow differences caused by pilling. These processes ultimately generate image feature data containing the visual characteristics of the fabric surface.

[0035] The system then fuses the micro-thermal radiation feature data and image feature data. Feature concatenation connects the two types of feature vectors in a dimensional way, forming a higher-dimensional comprehensive feature vector. Since the dimensions and numerical ranges of different features may vary significantly, the system performs normalization, typically using min-max normalization or Z-score standardization to map each feature value to a uniform numerical range. This process ensures the fairness of weights for different features in the subsequent model, preventing certain features from dominating model training due to their larger values. The multimodal feature data generated through fusion includes both the thermophysical properties of the fabric surface and visual morphological characteristics, providing a more comprehensive information foundation for accurate rating.

[0036] Finally, the multimodal feature data is input into a neural network model for recognition and classification. A neural network model is a computational model inspired by the structure of neurons in the human brain, consisting of multiple processing layers. It can automatically learn the complex mapping relationship between input features and output labels through forward and backward propagation algorithms. During the training phase, the model learns how to determine the pilling level based on multimodal feature data using a large number of labeled fabric samples. In the application phase, the trained model can infer the features of new fabric samples and output the corresponding pilling level data. The non-linear processing capability of the neural network model enables it to effectively mine deep patterns in multimodal features, thereby achieving accurate rating.

[0037] This method combines thermophysical detection and visual imaging techniques to achieve multi-dimensional characterization of fabric surface conditions. Thermal infrared polarization imaging provides deep physical characteristics that are difficult to obtain using traditional visual methods, while visible light imaging retains intuitive surface morphology information. By fusing these two types of information, the system can assess fabric condition from more perspectives, significantly improving the comprehensiveness and accuracy of the rating. The introduction of a neural network model enables the system to automatically learn complex feature-rating relationships, avoiding the limitations of manually designed judgment rules. The entire process, from data acquisition and feature extraction to feature fusion and intelligent recognition, forms a complete automated rating system, effectively overcoming the subjectivity and instability of manual rating and providing a reliable technical means for fabric quality inspection.

[0038] In some implementations, S1 includes:

[0039] S11: Based on the fabric sample, apply thermal excitation to the fabric sample to generate a thermally excited fabric sample;

[0040] S12: Based on the thermally excited fabric sample, thermal infrared images at different polarization angles are acquired using a thermal infrared polarization camera to generate multiple thermal infrared image data.

[0041] S13: Based on multiple thermal infrared image data, combine to generate thermal infrared polarization image sequence data.

[0042] In the acquisition of thermal infrared polarization image sequence data, thermal excitation of the fabric sample is the first step. Thermal excitation is a process of briefly and uniformly heating the fabric surface using a controllable heat source; common methods include halogen lamp irradiation and hot air gun blowing. The purpose of this step is to induce a small, controllable increase in the fabric surface temperature, thereby stimulating its thermal response characteristics. After thermal excitation, the surface temperature distribution of the fabric sample will exhibit a specific pattern due to differences in material thermal properties and structural characteristics, providing the necessary initial conditions for subsequent thermal attenuation analysis.

[0043] Then, images of the thermally excited fabric samples were acquired using a thermal infrared polarization camera. During acquisition, the camera obtained thermal infrared images at multiple different polarization angles by rotating the polarization filter or using amplitude-splitting polarization imaging. For example, thermal images could be acquired at four typical polarization directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees. Images were acquired at different polarization angles because the fabric surface and its pilling structure respond differently to thermal radiation in different polarization directions. This polarization characteristic is closely related to the orientation and regularity of the surface microstructure. Multi-angle polarization imaging allows for the acquisition of more comprehensive information on surface scattering characteristics.

[0044] Finally, the system combines multiple thermal infrared image data acquired at different polarization angles in a specific order and format to generate a complete thermal infrared polarization image sequence. This dataset not only contains spatial temperature distribution information but also polarization characteristic variations, forming a three-dimensional dataset rich in physical information. This data organization facilitates subsequent Stokes parameter calculations and polarization characteristic analysis, laying a data foundation for the extraction of micro-thermal radiation characteristics.

[0045] In some implementations, S2 includes:

[0046] S21: Calculate the polarization degree of each pixel based on thermal infrared polarization image sequence data;

[0047] S22: Based on thermal infrared polarization image sequence data, analyze the thermal decay curve and calculate the thermal decay time constant of each pixel;

[0048] S23: Combine polarization degree and thermal decay time constant to generate micro-thermal radiation characteristic data.

[0049] In the micro-thermal radiation feature extraction stage, the polarization degree of each pixel is first calculated based on the thermal infrared polarization image sequence data. The calculation of the polarization degree requires the use of Stokes parameters, which are obtained by analyzing the intensity images at different polarization angles. Specifically, for each pixel location, the system extracts its intensity value at each polarization angle, and then calculates the four parameters I, Q, U, and V according to the Stokes parameter calculation formula. The polarization degree is then calculated using a specific relationship between the Q, U, and V parameters and I. This calculation process quantifies the degree of polarization of thermal radiation at each pixel, reflecting the modulation effect of surface microstructure on thermal radiation.

[0050] Simultaneously, the system also calculates the thermal decay time constant for each pixel based on the same set of thermal infrared polarization image sequence data. The calculation process first requires extracting the temperature value of each pixel over time, forming a thermal decay curve of temperature change over time. Then, a curve fitting algorithm (such as the least squares method) is used to fit the temperature decay data to an exponential decay model, and the characteristic time constant is extracted from the fitting parameters. This time constant characterizes the rate of temperature decay at that point, reflecting the thermal properties of the local region.

[0051] Finally, the system performs feature-level fusion of the calculated polarization degree and thermal decay time constant. Fusion typically employs feature vector concatenation, combining these two parameters for each pixel into a two-dimensional feature vector. If necessary, statistical characteristics of these parameters (such as mean and variance) can be further extracted as overall features of the fabric sample. The micro-thermal radiation feature data generated through this fusion method simultaneously includes surface scattering properties and thermal property information, providing sufficient physical characteristic basis for subsequent rating analysis.

[0052] In some implementations, S21 includes:

[0053] S211: Calculate Stokes parameters based on thermal infrared polarization image sequence data;

[0054] S212: Calculate the polarization degree of each pixel based on Stokes parameters.

[0055] In the specific implementation of polarization degree calculation, the Stokes parameters first need to be calculated based on the thermal infrared polarization image sequence data. The Stokes parameters are a set of four parameters (I, Q, U, V), which together comprehensively describe the intensity and polarization state of an electromagnetic radiation beam. Parameter I represents the total intensity; Q represents the intensity difference between the horizontal and vertical polarization components; U represents the intensity difference between the +45° and -45° polarization components; and V represents the intensity difference between the right-hand and left-hand circular polarization components. For linear polarization thermal imaging systems, typically only parameters I, Q, and U need to be calculated.

[0056] The specific process for calculating Stokes parameters involves using the intensity values ​​of thermal infrared images acquired at different polarization angles and solving for them using specific linear combination formulas. For example, using intensity images I(0°), I(45°), and I(90°) at three polarization directions (0°, 45°, and 90°), the calculations can be performed using the formulas I = I(0°) + I(90°), Q = I(0°) - I(90°), and U = 2I(45°) - I(0°) - I(90°). The derivation of these formulas is based on the fundamental principles of polarization optics, ensuring the physical accuracy of the parameter calculations.

[0057] After obtaining the Stokes parameters, the degree of polarization (DOP) of each pixel can be calculated. The degree of polarization (DOP) is defined as the ratio of the intensity of the polarized portion to the total intensity, and the specific calculation formula is as follows: For linearly polarized systems, the V parameter is usually zero, and the formula can be simplified to... This calculation process converts the Stokes parameter into a scalar value, facilitating subsequent feature analysis and processing. This calculation method, based on a rigorous physical model, ensures the accuracy and reliability of the polarization characteristics.

[0058] In some implementations, S22 includes:

[0059] S221: Based on thermal infrared polarization image sequence data, extract data on temperature changes over time and generate temperature decay curve data;

[0060] S222: Calculate the thermal decay time constant for each pixel based on temperature decay curve data.

[0061] The first step in calculating the thermal decay time constant is to extract temperature-time data from thermal infrared polarization image sequence data. Since the thermal infrared image sequence is acquired at fixed time intervals after thermal excitation, the intensity value change of each pixel in the sequence directly reflects the temperature decay process at that point over time. The system establishes a time-temperature correspondence for each pixel, converting the intensity values ​​at different times in the sequence into temperature values, forming temperature decay curve data. This conversion process requires converting the digital intensity values ​​into actual temperature values ​​based on the camera's radiometric calibration parameters.

[0062] After obtaining the temperature decay curve data, the system calculates the thermal decay time constant for each pixel using curve fitting. A commonly used fitting model is the exponential decay model: T(t) = T0 + ΔT·exp(-t / τ), where T(t) is the temperature at time t, T0 is the ambient temperature, ΔT is the initial temperature rise, and τ is the thermal decay time constant. The system uses a nonlinear least squares fitting algorithm to fit the experimentally measured temperature-time data into this model, thereby solving for the optimal τ value. This fitting process ensures the accuracy and reliability of the time constant calculation.

[0063] The physical significance of the thermal decay time constant lies in its reflection of the material's thermal diffusion characteristics. A larger time constant indicates slower temperature decay at that point, signifying better thermal insulation performance; a smaller time constant indicates faster heat loss. In fabric pilling rating, pilling areas, due to their loose fibers and numerous air gaps, have higher thermal resistance, and their thermal decay time constant is typically greater than that of smooth areas. Therefore, this characteristic can effectively distinguish different states of the fabric surface, providing an important basis for automatic rating. By systematically analyzing the thermal decay characteristics of each pixel, a quantitative assessment of the degree of fabric pilling can be achieved.

[0064] In some implementations, S6 includes:

[0065] S61: Based on multimodal feature data, construct feature maps and generate graph structure data;

[0066] S62: Based on graph structure data, initial rating data is generated through graph neural network model identification;

[0067] S63: Based on the initial rating data, weighted rating data is generated through an attention mechanism.

[0068] S64: Generate pilling and fuzzing level data based on weighted rating data.

[0069] The process of constructing feature maps based on multimodal feature data is a crucial step in transforming high-dimensional feature vectors into graph-structured data. Graph-structured data is a non-Euclidean data structure composed of nodes and edges, effectively representing complex relationships between features. In practice, the system maps each feature vector to an independent node in the graph, with the node's attributes being the numerical values ​​of that feature vector. This mapping preserves the integrity of the original features while laying the foundation for establishing relationships between features. After generating the node set data, the system needs to calculate the similarity between feature vectors to construct the edge set data. Cosine similarity algorithm can be used for similarity calculation. This algorithm assesses the directional consistency of two vectors by calculating the cosine of the angle between them in space; a value closer to 1 indicates higher similarity. When the similarity exceeds a preset threshold, a connection edge is established between the corresponding nodes, and the edge weight is usually proportional to the similarity value. The final generated graph-structured data fully preserves the attribute information of the feature vectors and their interrelationships, providing a suitable input format for subsequent graph neural network processing.

[0070] When processing graph-structured data, graph neural network models first achieve feature propagation and aggregation through graph convolutional layers. Graph convolution is a convolutional operation that operates on graph structures. Its core idea is to generate new node representations by aggregating the features of a node itself and the features of its neighboring nodes. In specific implementation, each node collects feature information from its directly connected neighbors, performs a weighted sum based on edge weights, and then transforms it through a non-linear activation function. This process allows each node to perceive the structural features of its local neighborhood, thereby achieving deep feature fusion. After multiple layers of graph convolution processing, the node features retain both the original information and incorporate rich contextual information, generating convolutional feature data.

[0071] After obtaining the convolutional feature data, the system performs feature aggregation through pooling layers. Graph pooling aims to compress a variable number of node features into a fixed-dimensional global feature representation. Global average pooling is a common method, which averages the features of all nodes across all feature dimensions to generate a vector representing the overall graph features. Another method is global max pooling, which extracts the maximum value in each feature dimension, highlighting the most salient features. The pooled feature data integrates information from all nodes in the graph, providing a core feature representation for the final classification decision.

[0072] The introduction of an attention mechanism further enhances the model's feature selection capability. The system calculates the attention score for each node based on initial rating data, a process implemented through a trainable attention network. This network takes node features as input and calculates a scalar score representing node importance through fully connected layers and activation functions. Subsequently, a softmax function is used to normalize the attention scores of all nodes, ensuring their sum is 1; these normalized scores represent the weights of each node in the final decision. Based on these weights, the system performs a weighted summation of node features, generating weighted sum data. This adaptive weighting process dynamically focuses on the feature regions most important to the classification task, effectively suppressing noise interference. Finally, the weighted sum data is transformed into weighted rating data through fully connected layers and a softmax classifier, outputting the predicted probabilities for each level.

[0073] This method effectively organizes the relationships between multimodal features through a graph structure, enabling explicit expression of semantic associations between features. Graph convolution operations fully utilize the graph's topology, achieving effective propagation of feature information between neighborhoods and enhancing the feature representation capability of each node. Pooling layers generate compact graph-level representations by aggregating global information, providing comprehensive feature basis for hierarchical classification. The attention mechanism empowers the model to dynamically adjust feature weights, allowing it to focus on key information based on specific sample characteristics. These technical components work together to improve the efficiency of feature utilization and the accuracy of rating decisions, achieving a reliable conversion from multimodal features to the final rating result.

[0074] In some implementations, S61 includes:

[0075] S611: Based on multimodal feature data, each feature vector is mapped to a node to generate a set of node data;

[0076] S612: Based on node set data, calculate the similarity between feature vectors and construct edge set data;

[0077] S613: Generate graph structure data based on node set data and edge set data.

[0078] In constructing graph-structured data, mapping each feature vector to a node is the first step. Feature vectors originate from multimodal feature data, potentially corresponding to different spatial locations or feature modalities of the fabric sample. The system assigns a unique node identifier to each feature vector and uses its value as the node's feature attribute. This process preserves all original feature information while organizing the features into a set of nodes. The formation of this set of nodes establishes a one-to-one correspondence between features and graph nodes, laying the foundation for subsequently constructing a network of relationships between features.

[0079] Constructing an edge set requires calculating the similarity between feature vectors. Similarity calculation is crucial for establishing connections between nodes. Cosine similarity is a commonly used method, assessing similarity by measuring the difference in direction between two vectors. Specifically, the system divides the dot product of two feature vectors by the product of their magnitudes, obtaining a similarity value between -1 and 1. To establish effective connections, the system sets a similarity threshold; only when the calculated similarity exceeds this threshold is an edge connection established between the corresponding nodes. The edge weight is typically set to the similarity value, thus representing both the existence and strength of the connection. By traversing all node pairs and performing the above operations, a complete edge set is ultimately formed.

[0080] After obtaining the node set and edge set data, the system integrates them to generate complete graph structure data. Graph structure data is typically represented in the form of an adjacency matrix and a feature matrix. The adjacency matrix records the connection relationships between nodes and edge weights, while the feature matrix stores the feature vectors of all nodes. This data organization not only preserves the original feature information but also clearly defines the relationships between features, providing a structured input for graph neural networks. The generation of graph structure data allows the relationships between features to be explicitly expressed. This expression method better reflects the inherent connections between features and helps improve the efficiency of subsequent feature learning.

[0081] In some implementations, S62 includes:

[0082] S621: Based on graph structure data, convolutional feature data is generated through graph convolutional layers;

[0083] S622: Based on the convolutional feature data, pooling layers are used to aggregate the features to generate pooled feature data;

[0084] S623: Generate initial rating data based on the pooled feature data.

[0085] When processing graph-structured data, graph convolutional layers enhance features through neighborhood information aggregation. The core of graph convolution operations is a feature propagation mechanism defined on the graph structure, where each node updates its representation by aggregating features from its direct neighbors. Specifically, the system first determines the neighbor set for each node based on the adjacency matrix, then weights and sums the neighbor features according to edge weights. The weighted neighbor features are then linearly combined with the node's own features, and further processed by a non-linear activation function such as ReLU to generate updated node features. This process allows each node to perceive the feature distribution of its local neighborhood, thus achieving contextual enhancement of features. Through multi-layer graph convolution stacking, nodes can receive information from multi-hop neighbors, achieving deeper feature fusion and generating convolutional feature data.

[0086] The role of pooling layers is to aggregate the convolutional feature data into a graph-level representation. Since the number of nodes in the graph may vary, while the final classification requires a fixed-dimensional feature vector, pooling operations are needed to achieve this transformation. Global average pooling is a commonly used method; it averages the values ​​of all nodes across each feature dimension, generating a global vector with the same feature dimensions. This method treats the contributions of all nodes equally and is suitable for scenarios with relatively uniform feature distribution. Global max pooling is another option; it extracts the maximum value across each feature dimension, highlighting the most significant feature responses. Pooling operations effectively elevate node-level features to graph-level features, generating pooled feature data that contains comprehensive feature information from the entire graph.

[0087] The generation of initial rating data based on pooled feature data is achieved through a classification network. The pooled feature data, as a global feature representation, is input into a classifier composed of fully connected layers. These fully connected layers learn the complex mapping relationship between features and ratings through linear transformations and non-linear activation functions. The final stage typically uses a softmax function to transform the output into a probability distribution, representing the likelihood of a sample belonging to each rating. This probability distribution constitutes the initial rating data, providing preliminary classification results based on global graph features. The entire processing flow, from local feature aggregation to global feature extraction, and finally to the classification decision, forms a complete graph neural network processing chain.

[0088] In some implementations, S63 includes:

[0089] S631: Calculate the attention score for each node based on the initial rating data;

[0090] S632: Based on attention scores, perform weighted summation on the nodes to generate weighted summation data;

[0091] S633: Generate weighted rating data based on weighted summation data.

[0092] Attention mechanisms enhance the discriminative power of feature representations by adaptively weighting node features. Calculating attention scores is the core step of the attention mechanism, aiming to evaluate the importance of each node to the final classification task. The system typically employs a neural network-based attention calculation method, inputting node features into a fully connected layer and generating raw attention scores through an activation function. The weights of the fully connected layer are learned during training, enabling it to identify discriminative feature patterns for the classification task. To ensure comparability of scores between different nodes, the system uses a softmax function to normalize the raw scores of all nodes, ensuring that the sum of the normalized attention scores is 1.

[0093] Weighted summation based on attention scores is a crucial step in feature selection. The system multiplies the features of each node by its corresponding normalized attention score, and then sums all the weighted features to generate a weighted sum. This process essentially selectively enhances node features; highly important node features are amplified, while less important features are suppressed. The weighted sum is a global representation that integrates information from all nodes but emphasizes key features. It better reflects the feature components most important for the classification task than simple average pooling or max pooling.

[0094] The process of generating the final weighted rating data is similar to the initial rating, but it's based on attention-weighted feature representations. The weighted summation data is input into the classification network, where it's transformed into a rank probability distribution through fully connected layers and a softmax function. Because the input features have been refined by the attention mechanism, removing some noise and highlighting key features, the weighted rating data generated based on it typically has higher accuracy. The attention mechanism allows the model to dynamically adjust feature weights based on the characteristics of specific samples; this adaptive capability significantly improves the model's feature utilization efficiency and classification performance.

[0095] Example 2

[0096] like Figure 2 As shown, in a second aspect, the present invention provides an automatic fabric pilling and fuzzing rating device, the device employing an automatic fabric pilling and fuzzing rating method provided in any of the above embodiments, the device comprising:

[0097] The thermal infrared image acquisition module is used to acquire thermal infrared polarization image sequences of fabric samples using a thermal infrared polarization camera, and generate thermal infrared polarization image sequence data.

[0098] The micro-thermal radiation feature extraction module is used to extract the micro-thermal radiation features of fabric samples based on thermal infrared polarization image sequence data by calculating the degree of polarization and thermal decay time constant, and generate micro-thermal radiation feature data.

[0099] The visible light image acquisition module is used to acquire visible light images of fabric samples using a visible light camera, and generate visible light image data.

[0100] The image feature extraction module is used to extract image features of fabric samples based on visible light image data through grayscale processing, texture feature calculation, and color feature calculation, and generate image feature data.

[0101] The multimodal feature fusion module is used to generate multimodal feature data by fusing micro-thermal radiation feature data and image feature data through feature splicing and normalization processing.

[0102] The pilling / fuzzing grade determination module is used to identify and determine the pilling / fuzzing grade data of fabric samples based on multimodal feature data and a neural network model, including:

[0103] Based on multimodal feature data, feature maps are constructed to generate graph structure data;

[0104] Based on graph-structured data, initial rating data is generated through graph neural network model identification.

[0105] Based on the initial rating data, weighted rating data is generated by weighting through an attention mechanism;

[0106] Based on the weighted rating data, pilling and fuzzing level data are generated.

[0107] This device corresponds to the method provided in Embodiment 1, and will not be described in detail here.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method of automatically rating fabric pilling, characterized by, The method comprises the following steps: S1: based on the fabric sample, using a thermal infrared polarization camera to collect a thermal infrared polarization image sequence of the fabric sample, generating thermal infrared polarization image sequence data; S2: based on the thermal infrared polarization image sequence data, by calculating the degree of polarization and the thermal decay time constant, extracting the micro-thermal radiation features of the fabric sample, generating micro-thermal radiation feature data; S3: based on the fabric sample, using a visible light camera to collect a visible light image of the fabric sample, generating visible light image data; S4: based on the visible light image data, by grayscale processing, texture feature calculation and color feature calculation, extracting the image features of the fabric sample, generating image feature data; S5: based on the micro-thermal radiation feature data and the image feature data, by feature splicing and normalization processing, fusion generates multi-modal feature data; S6: based on the multi-modal feature data, through a neural network model to identify and determine the pilling grade data of the fabric sample; wherein, S6 comprises: S61: based on the multi-modal feature data, constructing a feature map, generating graph structure data; S62: based on the graph structure data, identifying through a graph neural network model, generating initial rating data; S63: based on the initial rating data, generating weighted rating data through attention mechanism weighting; S64: based on the weighted rating data, generating the pilling grade data.

2. A method of automatically grading a fabric pilling according to claim 1, wherein, S1 comprises: S11: based on the fabric sample, applying thermal excitation to the fabric sample to generate a fabric sample after thermal excitation; S12: based on the fabric sample after thermal excitation, using a thermal infrared polarization camera to collect thermal infrared images under different polarization angles, generating multiple thermal infrared image data; S13: based on the multiple thermal infrared image data, combining to generate thermal infrared polarization image sequence data.

3. A method of automatically grading a fabric pilling according to claim 1, wherein S2 Comprise: S21: based on the thermal infrared polarization image sequence data, the degree of polarization of each pixel point is calculated; S22: based on the thermal infrared polarization image sequence data, the thermal decay time constant of each pixel point is calculated by analyzing the thermal decay curve; S23: fusion of the degree of polarization and the thermal decay time constant, generating micro-thermal radiation feature data.

4. A method of automatically grading the pilling of a fabric as defined in claim 3, wherein S21 comprises: S211: based on the thermal infrared polarization image sequence data, calculating Stokes parameters; S212: based on the Stokes parameters, calculating the degree of polarization of each pixel point.

5. A method of automatically grading the pilling of a fabric as defined in claim 3, wherein, S22 comprises: S221: based on the thermal infrared polarization image sequence data, extracting the data of temperature change with time, generating temperature decay curve data; S222: based on the temperature decay curve data, calculating the thermal decay time constant of each pixel point.

6. A method of automatically grading the pilling of a fabric as defined in claim 1, wherein S61 comprises: S611: based on the multi-modal feature data, mapping each feature vector to a node to generate node set data; S612: based on the node set data, calculating the similarity between feature vectors, constructing edge set data; S613: based on the node set data and the edge set data, generating graph structure data.

7. A method of automatically grading the pilling of a fabric as defined in claim 1, wherein S62 comprises: S621: based on the graph structure data, generating convolution feature data through a graph convolution layer; S622: based on the convolution feature data, generating pooled feature data through a pooling layer aggregation; S623: based on the pooled feature data, generating initial rating data.

8. A method of automatically grading the pilling of a fabric as defined in claim 1, wherein S63 comprises: S631: Calculate the attention score of each node based on the initial rating data; S632: Based on the attention score, weighted sum of nodes is generated, and weighted sum data is generated; S633: Based on the weighted sum data, generate weighted rating data.

9. An apparatus for automatically rating the pilling of a fabric, characterized in that, The device adopts the automatic fabric pilling and fuzz rating method of any one of claims 1 to 8, and the device comprises: a thermal infrared image acquisition module for acquiring a thermal infrared polarized image sequence of the fabric sample based on the fabric sample using a thermal infrared polarized camera, and generating thermal infrared polarized image sequence data; a micro-thermal radiation feature extraction module for extracting micro-thermal radiation features of the fabric sample by calculating the degree of polarization and the thermal decay time constant based on the thermal infrared polarized image sequence data, and generating micro-thermal radiation feature data; a visible light image acquisition module for acquiring a visible light image of the fabric sample based on the fabric sample using a visible light camera, and generating visible light image data; an image feature extraction module for extracting image features of the fabric sample by grayscale processing, texture feature calculation and color feature calculation based on the visible light image data, and generating image feature data; a multi-modal feature fusion module for fusing and generating multi-modal feature data by feature splicing and normalization processing based on the micro-thermal radiation feature data and the image feature data; a grade determination module for identifying and determining the pilling and fuzz grade data of the fabric sample based on the multi-modal feature data through a neural network model, comprising: based on the multi-modal feature data, constructing a feature map to generate graph structure data; based on the graph structure data, identifying through a graph neural network model to generate initial rating data; based on the initial rating data, generating weighted rating data through attention mechanism weighting; based on the weighted rating data, generating pilling and fuzz grade data.

Images