A cloth material identification method and system based on label relationship

Abstract

The application belongs to the technical field of cloth identification, and discloses a cloth material quality identification method and system based on label relations, which comprises the following steps: obtaining a sample set to be measured, semantic features, original classification scores, a label vector of cloth, and constructing an initial relation graph; inputting the initial relation graph into a first graph attention network to obtain shallow relation representation; connecting the semantic features with the shallow relation representation to obtain joint features of each node; inputting the joint features into a second graph attention network to form label encoding; using the label encoding to map the original classification scores into material classification scores corresponding to the sample material quality; inputting the shallow relation representation into a third graph attention network to obtain deep relation representation, and extracting a label compensation vector; using the label compensation vector to compensate the material classification scores to obtain target classification scores, and comparing the target classification scores with positive and negative sample thresholds to determine material information. The application considers the connectivity between different cloths, and significantly improves the identification accuracy of cloth material quality.

Description

Technical Field

[0001] This invention belongs to the technical field of fabric recognition, and more specifically, relates to a method and system for identifying fabric material based on tag relationships. Background Technology

[0002] Fabrics play a vital role in all aspects of human life, and identifying the internal fiber composition of fabrics is beneficial for understanding their properties, comparing them, and recycling them. Therefore, fabric material identification has wide applications in manufacturing, fashion design, home furnishings, and forensic medicine.

[0003] Methods for identifying fabric materials can be divided into two categories: traditional methods and intelligent methods. Traditional methods mainly include physical and chemical methods, near-infrared spectroscopy, etc. These methods require long detection cycles, rely on expert experience and expensive and complex equipment, and are often destructive to the fabric. Therefore, these methods lack universality. With the rapid development of artificial intelligence technology, intelligent fabric material identification methods can learn the texture features and deeper semantic features from massive amounts of fabric image data to achieve accurate material identification. Currently, intelligent identification methods such as deep learning are becoming a development trend, effectively overcoming the limitations of traditional methods and offering greater advantages.

[0004] A fabric often contains multiple fiber materials, making fabric material identification a multi-label recognition task. Currently, many deep learning methods in fabric material identification transform the problem into multiple binary classification problems or a single multi-class classification problem for easier solution. However, these methods do not consider the relationships between different fiber material labels. These relationships are ubiquitous; for example, fabrics made of cotton and polyester are common, while leather is rarely found in the same fabric. Therefore, there is a pressing need to design a fabric material identification method based on label relationships to improve the accuracy of fabric material identification. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the prior art, the present invention provides a method and system for identifying fabric materials based on label relationships. This application takes into account the connection between different fabrics, which significantly improves the accuracy of fabric material identification.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for identifying fabric material based on label relationships is provided, comprising: S1: acquiring surface images of the fabric to be tested to obtain a sample set; S2: inputting the sample set into a deep convolutional neural network model to obtain semantic features and original classification scores of each sample in the sample set; S3: obtaining label vectors for each type of fabric in the sample set, and using the label vectors for each type of fabric as initial node features to construct an initial relationship graph; S4: inputting the initial relationship graph into a first graph attention network to obtain shallow relationship representations of each node; S5: performing a connection operation between the semantic features and the shallow relationship representations of each node to perform a specific... S6: The joint features of each node are obtained by feature fusion; S7: The joint features are input into the second graph attention network and the node features are updated. The updated node features constitute the label encoding; S8: The original classification score is mapped to a material classification score that corresponds one-to-one with the sample material using the label encoding; S9: The shallow relation representation of each node is input into the third graph attention network to obtain the deep relation representation of each node, and the deep relation representation corresponding to the material category with the highest material classification score is extracted as the label compensation vector; S10: The material classification score is compensated using the label compensation vector to obtain the target classification score. The target classification score is compared with the positive and negative sample thresholds to determine the material information.

[0007] Preferably, step S2 specifically involves: performing layer-by-layer convolution on the sample to be tested using multiple convolutional layers in the deep convolutional neural network model, and inputting the final output feature map into a global average pooling layer to obtain one-dimensional semantic features; and inputting the semantic features into a linear classifier to obtain the original classification score for each sample.

[0008] Preferably, step S3 specifically involves: using an embedding layer to map the label words of each type of fabric in the test sample set into label vectors, so that each type of fabric material label uniquely obtains a label vector.

[0009] Preferably, step S4 specifically involves: for the initial features of nodes in the relationship graph, using the graph attention mechanism in the first graph attention network to calculate the attention coefficients of the neighboring nodes for each node, so as to model the shallow global connections between different materials, and updating the node features. The updated node features constitute the shallow relationship representation.

[0010] Preferably, before performing the connection operation in step S5, the method further includes transforming the length of the semantic features through a linear layer to be consistent with the length of each shallow relation representation, and then connecting them according to the same feature length.

[0011] Preferably, step S6 specifically involves: for the joint features of each node, using the graph attention mechanism in the second graph attention network to calculate the attention coefficients of each node's neighboring nodes, and updating the node features. The updated node features constitute the label encoding. Since the joint features not only contain global label information beyond the image, but also semantic information (local) within the image, the graph attention mechanism enables a deep fusion of the two in relational modeling. Therefore, the label encoding implies the connection between the global and local aspects.

[0012] Preferably, step S8 specifically involves: for the shallow relationship representation of each node, using the graph attention mechanism in the third graph attention network to calculate the attention coefficients of the neighboring nodes of each node, so as to model the deep global connection between different nodes, and updating the node features. The updated node features constitute the deep relationship representation.

[0013] Preferably, the step S9 of using the label compensation vector to compensate the material classification score specifically involves superimposing the label compensation vector with the material classification score to obtain the target classification score.

[0014] This application, in another aspect, provides a fabric material recognition system based on label relationships, comprising: an image sampling module for acquiring surface images of the fabric to be tested to obtain a sample set; an image feature extraction module for inputting the samples to be tested into a deep convolutional neural network model to obtain semantic features of each sample in the sample set and the original classification score of each sample; a graph construction module for obtaining label vectors of each type of fabric in the sample set and constructing an initial relationship graph using the label vectors of each type of fabric as initial node features; a first global relationship modeling module for inputting the initial relationship graph into a first graph attention network to obtain shallow relationship representations of each node; and a feature fusion module for connecting the semantic features with the shallow relationship representations of each node to perform feature fusion. The system obtains the joint features of each node; the joint relationship modeling module is used to input the joint features into the second graph attention network and update the node features, and the updated node features constitute the label encoding; the label encoding mapping module is used to map the original classification score into a material classification score that corresponds one-to-one with the sample material using the label encoding; the second global relationship modeling module is used to input the shallow relationship representation of each node into the third graph attention network to obtain the deep relationship representation of each node, and extract the deep relationship representation corresponding to the material category with the highest material classification score as the label compensation vector; the recognition module is used to compensate the material classification score using the label compensation vector to obtain the target classification score, and compare the target classification score with the positive and negative sample thresholds to determine the material information.

[0015] In summary, compared with the prior art, the fabric material identification method and system based on tag relationships provided by the present invention have the following advantages:

[0016] 1. This application employs a deep convolutional neural network to extract features from the test samples, then performs relationship modeling, considering the global and local connections between different material labels. After label compensation, the global relationship information of the material classification scores is further enhanced, resulting in a more complete recognition result. Compared to existing intelligent material recognition methods, this invention considers the connections between different material labels, enabling a more comprehensive consideration of the co-occurrence possibility and mutual independence of various materials, significantly improving the accuracy of fabric material recognition.

[0017] 2. This invention first utilizes a deep convolutional neural network model in the image feature extraction module to extract semantic features from the original surface image of the fabric under test. Then, through a first global relation modeling module, a second global relation modeling module, and a joint relation modeling module, the device is endowed with relation modeling capabilities, considering the global and local connections between different material labels. In the image feature extraction module, the output raw classification score only contains local information within the image; however, after passing through the label encoding and mapping module, the output material classification score not only contains local feature information but also includes joint global and local relational information; after passing through the label compensation module, the global relational information of the material classification score is further enhanced, resulting in a more complete recognition result. Compared to existing intelligent material recognition systems, this invention considers the connections between different material labels, and can more fully consider the co-occurrence possibility and mutual independence of various fiber materials, thereby improving the accuracy of fabric material recognition. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the steps of a fabric material identification method based on tag relationships;

[0019] Figure 2 This is a schematic diagram of a fabric material identification system based on tag relationships;

[0020] Figure 3 This is a schematic diagram illustrating the structure of the initial relationship graph;

[0021] Figure 4 This is a schematic diagram of the execution flow of a fabric material recognition system based on tag relationships. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0023] This invention provides a method for identifying fabric material based on tag relationships, such as... Figure 1 As shown, the method includes the following steps S1 to S9.

[0024] S1: Acquire surface images of the fabric to be tested to obtain a sample set for testing.

[0025] During the training phase, a sufficient number of fabric types must be tested. In this embodiment, the test sample set includes ten different fabric materials: wool, acrylic, linen, silk, polyester, viscose, spandex, cotton, leather, and nylon.

[0026] The image resolution is set to 224×224.

[0027] S2: Input the test samples into a deep convolutional neural network model to obtain the semantic features of each sample in the test sample set and the original classification score of each sample.

[0028] In a further preferred embodiment, the test sample is convolved layer by layer using multiple convolutional layers in the deep convolutional neural network model, and the final output feature map is input into a global average pooling layer to obtain one-dimensional semantic features; the semantic features are then input into a linear classifier to obtain the original classification score for each sample.

[0029] Specifically, multiple convolutional layers in a deep convolutional neural network model are used to extract features from the test sample from shallow to deep. After feature extraction by stacked convolutional and pooling layers, 768 feature maps of size 7×7 are obtained for each test sample.

[0030] The feature map is processed by a global average pooling layer to obtain image semantic features with a dimension of 1×768. These features contain abstract material semantic information learned from the image and belong to local information.

[0031] The linear classifier is a fully connected layer that maps semantic features of dimension 1×768 into raw classification scores of dimension 1×1000.

[0032] S3: Obtain the label vector of each fabric in the sample set to be tested, and use the label vector of each fabric as the initial node feature to construct the initial relationship graph.

[0033] like Figure 3As shown, the labels corresponding to the fabric materials are used to obtain label vectors through an embedding layer. The embedding layer maps label words to one-dimensional feature vectors representing the words. Through the embedding layer, each fabric material label uniquely obtains a label vector, meaning the number of label vectors equals the number of fabric fiber material categories.

[0034] In this embodiment, all fabric material tags are mapped to tag vectors of dimension 1×128 through the embedding layer. Each material tag uniquely obtains a tag vector, and ten material categories correspond to ten tag vectors, forming a total of 10×128 tag vectors.

[0035] This step involves constructing a graph by treating each material label as a node and initializing the node features using label vectors. Specifically, the label vector corresponding to each material category is used as the initial node features to construct the graph, meaning the number of nodes in the graph equals the number of fiber material categories of the fabric, and all nodes in the graph are connected by edges.

[0036] In this embodiment, the label vector corresponding to each material category is used as the initial node feature to construct the graph. Therefore, the graph contains 10 nodes, each with a feature dimension of 1×128. Each node represents a certain type of material label, and the labels represented by each node are different. All nodes in the graph are connected to each other by edges, so there are a total of 45 edges across the 10 nodes.

[0037] S4: Input the initial relationship graph into the first graph attention network to obtain shallow relationship representations of each node.

[0038] The initial relationship graph is implicitly modeled using a first graph attention network. Different attention coefficients are set for the neighboring nodes of different nodes through the graph attention mechanism, representing the correlation between different nodes, and the node features are updated accordingly.

[0039] In this embodiment, a multi-head graph attention mechanism is used to update node features. There are 4 heads in the multi-head graph. The dimension of the node features after each head is updated is 1×256. The four attention features are finally fused through a connection operation. Therefore, the dimension of the node features before the update is 1×128, and the dimension of the node features after the update is 1×1024.

[0040] The updated node features constitute the shallow relation representation. In the graph attention mechanism, different nodes calculate attentional connections with each other and update their own node features based on these connections. Since only one layer of the graph attention network is used, the shallow relation representation contains shallow global label connections. The number of shallow relation representations is equal to the number of fiber material categories in the fabric, corresponding one-to-one with each material category. Therefore, the overall dimension of all shallow relation representations is 10 × 1024. This step achieves shallow global information relationship modeling.

[0041] S5: Connect the semantic features with the shallow relationship representations of each node to perform feature fusion and obtain the joint features of each node.

[0042] The semantic features are linked with the shallow relation representations corresponding to each node to perform feature fusion, obtain joint features, and form new node features.

[0043] In a further preferred embodiment, between join operations, a linear layer transforms the length of the semantic features to be consistent with the length of each shallow relation representation, i.e., transforms it to 1×1024, and joins are performed according to the same feature length. The semantic features with the transformed lengths are then joined with the shallow relation representations corresponding to each node to perform feature fusion, thereby obtaining the joint features and forming new node features. The dimension of this joint feature is 10×2048.

[0044] S6: Input the joint features into the second graph attention network and update the node features. The updated node features constitute the label encoding.

[0045] The joint features are input into the second graph attention network for implicit relationship modeling. Different attention coefficients are set for the neighboring nodes of different nodes through the graph attention mechanism, representing the correlation between different nodes. The node features are updated accordingly, and the updated node features constitute the label encoding. Since the joint features not only contain global label information beyond the image, but also semantic information (local) within the image, the relationship modeling of the graph attention mechanism enables a deep fusion of the two. Therefore, the label encoding implies a fusion of global and local connections.

[0046] In this embodiment, a multi-head graph attention mechanism is used to update node features. There are four heads in the multi-head graph, and the dimension of the updated node features of each head is 1×1000. The four attention features are finally fused by averaging. Therefore, the dimension of the node features before the update is 1×2048, and the dimension of the updated node features is 1×1000.

[0047] S7: The original classification score is mapped to a material classification score that corresponds one-to-one with the sample material using the label encoding.

[0048] The number of label codes is equal to the number of fiber material categories of the fabric, and there is a one-to-one correspondence between the material categories. Therefore, the overall dimension of the label codes is 10×1000.

[0049] The original classification score is mapped to a material classification score using the label encoding. In this embodiment, a test sample can obtain one original classification score with a dimension of 1×1000. Simultaneously, a test sample can obtain 10 label codes, each corresponding to a material, with an overall dimension of 10×1000 and each label code having a dimension of 1×1000. Through matrix multiplication, the original classification score is multiplied by the transpose of the label code for one material class to obtain a classification score corresponding to that material class. The label codes for all 10 material classes are multiplied by the original classification scores to obtain 10 different material classification scores. Overall, the above process can be viewed as performing matrix multiplication of the original classification score (dimension 1×1000) with the overall label encoding device (dimension 1000×10) to obtain a material classification score with a dimension of 1×10.

[0050] S8: Input the shallow relation representations of each node into the third graph attention network to obtain the deep relation representations of each node, and extract the deep relation representations corresponding to the material category with the highest material classification score as the label compensation vector.

[0051] The shallow relationship representation of each node is implicitly modeled through the third graph attention network layer. Different attention coefficients are set for the neighboring nodes of different nodes through the graph attention mechanism, which represent the correlation between different nodes, and the node features are updated accordingly.

[0052] The node feature is updated using a multi-head graph attention mechanism with four heads. The dimension of the updated node feature is 1×10 for each head. The four attention features are then fused by averaging. Therefore, the dimension of the node feature before the update is 1×1024, and the dimension of the updated node feature is 1×10.

[0053] The updated node features constitute a deep relation representation. In the graph attention mechanism, different nodes will calculate attention connections with each other and update their own node features based on these attention connections. Since the shallow relation representation has already used one layer of graph attention network, and another layer of graph attention network is used, the deep relation representation contains shallow global label connections.

[0054] The number of deep relation representations is equal to the number of fiber material categories of the fabric, and there is a one-to-one correspondence with the material categories. Therefore, the overall dimension of the deep relation representation is 10×10.

[0055] Guided by the material classification score, the deep relation representation corresponding to the material category with the highest material classification score is extracted to form the label compensation vector. For example, if cotton has the highest material classification score among the ten materials "wool, acrylic, linen, silk, polyester, viscose, spandex, cotton, leather, and nylon", then the deep relation representation corresponding to the "cotton" material is extracted as the label compensation vector. Therefore, the dimension of the label compensation vector is 1×10.

[0056] S9: The material classification score is compensated using the label compensation vector to obtain the target classification score. The target classification score is then compared with the positive and negative sample thresholds to determine the material information.

[0057] In a further preferred embodiment, the label compensation vectors are added together to compensate for the material classification score. Since both the material classification score and the label compensation vector have a dimension of 1×10, the dimension of the material classification after label compensation remains 1×10. The compensated material score constitutes the final predicted score, which has a dimension of 1×10. The predicted score is normalized to the range [0,1] using a Sigmoid classifier. Each predicted score represents the predicted probability of each material in the test sample, thereby outputting the material identification result of the fabric to be tested.

[0058] For example, if the positive and negative sample threshold is set to 0.5, and the predicted target classification score is wool 0.8, leather 0.2, silk 0.1, and cotton 0.6, then the predicted scores of cotton and wool are greater than the positive and negative sample threshold, and the final prediction result is wool and cotton.

[0059] This application also provides a fabric material identification system based on tag relationships, such as... Figure 2 and Figure 4 As shown, the system includes an image sampling module 101, an image feature extraction module 102, a graph construction module 103, a first global relation modeling module 104, a feature fusion module 105, a joint relation modeling module 106, a label encoding and mapping module 107, a second global relation modeling module 108, and a recognition module 109, wherein:

[0060] Image sampling module 101: Used to perform step S1 above, acquire surface images of the fabric to be tested, and obtain a sample set to be tested. The sample set to be tested is then input into image feature extraction module 102.

[0061] Image feature extraction module 102: This module performs step S2 by inputting the test sample into a deep convolutional neural network model to obtain the semantic features of each sample in the test sample set and the original classification score of each sample. In this embodiment, the test sample is input into the deep convolutional neural network model to obtain the semantic features of the test sample, which have a dimension of 1×768, and the original classification score is obtained through a linear classifier, which has a dimension of 1×1000. The semantic features are subsequently input into the feature fusion module 105, and the original classification score is subsequently input into the label encoding mapping module 107.

[0062] Graph construction module 103: This module executes step S3 to obtain the label vector for each type of fabric in the test sample set, and uses the label vector for each type of fabric as the initial node feature to construct an initial relationship graph. In this embodiment, each material label is used as a node in the graph for construction. The initial number of nodes is 10, and the label vector is used to initialize the node features. The overall dimension of the features is 10×128. All 10 nodes in the graph are connected in pairs, with a total of 45 edges.

[0063] The first global relationship modeling module 104 is used to execute the above step S4, inputting the initial relationship graph into the first graph attention network to obtain shallow relationship representations of each node. In this embodiment, the initial node features are input into the first graph attention network layer to obtain the shallow relationship representations containing global label connections, with an overall dimension of 10×768; the shallow relationship representations are subsequently input into the feature fusion module 105 and the second global relationship construction model 108 respectively.

[0064] Feature fusion module 105: This module performs step S5 above, connecting the semantic features with the shallow relation representations of each node to achieve feature fusion and obtain joint features for each node. In this embodiment, the semantic features are fused with the shallow relation representations to obtain the joint features, which have an overall dimension of 10×2048. The joint features are subsequently input into the joint relation modeling module 106.

[0065] Joint Relationship Modeling Module 106: This module performs step S6, inputs the joint features into the second graph attention network, and updates the node features. The updated node features constitute the label encoding. In this embodiment, the joint features are input into the second graph attention network layer to obtain the label encoding that integrates local image information and global label information, and then the relationship modeling of the joint global and local information is performed. The dimension of the label encoding is 10×1000, which is subsequently input into the label encoding mapping module 107.

[0066] Label encoding mapping module 107: used to perform the above step S7, using the label encoding to map the original classification score into a material classification score that corresponds one-to-one with the sample material. In this embodiment, the label encoding is used to map the original classification score into a material classification score, with an overall dimension of 1×10; the material classification score is subsequently input to the recognition module 109.

[0067] The second global relationship modeling module 108 is used to execute the above step S8, inputting the shallow relationship representations of each node into the third graph attention network to obtain the deep relationship representations of each node. In this embodiment, the shallow relationship representations are input into the third graph attention network layer to obtain the deep relationship representations containing global label connections, with an overall dimension of 10×10. Under the guidance of the material classification score, the label compensation vector is obtained, with a dimension of 1×10. The deep relationship representation corresponding to the material category with the highest material classification score is extracted as the label compensation vector, which is subsequently input into the recognition module 109.

[0068] Identification module 109: is used to perform the above step S9, using the label compensation vector to compensate the material classification score to obtain the target classification score, and comparing the target classification score with the positive and negative sample thresholds to determine the material information.

[0069] This label-based fabric material recognition system has three different information perception modes:

[0070] (1) The image feature extraction module 102 senses local information and extracts features from the information inside the image;

[0071] (2) The first global relationship modeling module 104 and the second global relationship modeling module 108 sense global information and perform shallow and deep global relationship modeling on the label information beyond the image, respectively.

[0072] (3) The joint relationship modeling module 106 simultaneously senses global and local information and integrates the two to model the label relationship.

[0073] Therefore, this application considers the global and local connections between different material tags. Compared with existing intelligent material recognition technologies, this invention considers the connections between different material tags, and can more fully account for the co-occurrence possibility and mutual independence of various fiber materials, thereby improving the accuracy of fabric material recognition.

[0074] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying fabric material based on tag relationships, characterized in that, include: S1: Acquire surface images of the fabric to be tested to obtain the sample set to be tested; S2: Input the test samples into a deep convolutional neural network model to obtain the semantic features of each sample in the test sample set and the original classification score of each sample; S3: Obtain the label vector of each fabric in the sample set to be tested, and use the label vector of each fabric as the initial node feature to construct the initial relationship graph; S4: Input the initial relationship graph into the first graph attention network to obtain shallow relationship representations of each node; S5: The semantic features are linked with the shallow relationship representations of each node to perform feature fusion and obtain the joint features of each node. S6: Input the joint features into the second graph attention network and update the node features. The updated node features constitute the label encoding. S7: The original classification score is mapped to a material classification score that corresponds one-to-one with the sample material using the label encoding; S8: Input the shallow relation representations of each node into the third graph attention network to obtain the deep relation representations of each node, and extract the deep relation representations corresponding to the material category with the highest material classification score as the label compensation vector. S9: The material classification score is compensated using the label compensation vector to obtain the target classification score. The target classification score is then compared with the positive and negative sample thresholds to determine the material information.

2. The fabric material identification method according to claim 1, characterized in that, Step S2 is as follows: The test sample is convolved layer by layer using multiple convolutional layers in the deep convolutional neural network model, and the final output feature map is input into a global average pooling layer to obtain one-dimensional semantic features. The semantic features are input into a linear classifier to obtain the original classification score for each sample.

3. The fabric material identification method according to claim 1, characterized in that, Step S3 is as follows: The embedding layer is used to map the label words of each fabric in the test sample set into label vectors, and each fabric material label uniquely obtains a label vector.

4. The fabric material identification method according to any one of claims 1 to 3, characterized in that, Step S4 is as follows: For the initial features of nodes in the relationship graph, the graph attention mechanism in the first graph attention network is used to calculate the attention coefficients of the neighboring nodes of each node to model the shallow global connections between different materials, and to update the node features. The updated node features constitute the shallow relationship representation.

5. The fabric material identification method according to claim 1, characterized in that, Before performing the connection operation in step S5, the semantic features are transformed to be consistent with the length of each shallow relation representation through a linear layer, and then connected according to the same feature length.

6. The fabric material identification method according to any one of claims 1 to 3, characterized in that, Step S6 is as follows: For the joint features of each node, the graph attention mechanism in the second graph attention network is used to calculate the attention coefficients of its neighboring nodes for each node, and the node features are updated. The updated node features constitute the label encoding.

7. The fabric material identification method according to claim 1, characterized in that, Step S8 is as follows: For the shallow relationship representation of each node, the graph attention mechanism in the third graph attention network is used to calculate the attention coefficient of its neighboring nodes for each node, so as to model the deep global connection between different nodes and update the node features. The updated node features constitute the deep relationship representation.

8. The fabric material identification method according to claim 1, characterized in that, The step S9, which involves using the label compensation vector to compensate the material classification score, specifically involves: The target classification score is obtained by superimposing the label compensation vector with the material classification score.

9. A fabric material identification system based on tag relationships, characterized in that, include: Image sampling module: used to acquire surface images of the fabric to be tested and obtain a sample set to be tested; Image feature extraction module: used to input the test samples into a deep convolutional neural network model to obtain the semantic features of each sample in the test sample set and the original classification score of each sample; Graph construction module: used to obtain the label vector of each fabric in the sample set to be tested, and use the label vector of each fabric as the initial node feature to construct the initial relationship graph; First global relation modeling module: used to input the initial relation graph into the first graph attention network to obtain shallow relation representations of each node; Feature fusion module: used to connect the semantic features with the shallow relationship representations of each node to perform feature fusion and obtain the joint features of each node; Joint relationship modeling module: used to input the joint features into the second graph attention network and update the node features. The updated node features constitute the label encoding. Label encoding mapping module: used to map the original classification score into a material classification score that corresponds one-to-one with the sample material using the label encoding; The second global relationship modeling module is used to input the shallow relationship representations of each node into the third graph attention network to obtain the deep relationship representations of each node, and extract the deep relationship representation corresponding to the material category with the highest material classification score as the label compensation vector. The identification module is used to compensate the material classification score using the label compensation vector to obtain the target classification score, and compare the target classification score with the positive and negative sample thresholds to determine the material information.

Images