Flame-retardant material surface defect image recognition method, device and equipment and storage medium

Abstract

The invention relates to the technical field of image recognition, and discloses a flame-retardant material surface defect image recognition method, device and equipment and a storage medium, and the method comprises the steps: carrying out the multi-angle image collection and preprocessing of a flame-retardant material combustion test sample, and obtaining a standardized multi-view image data set; performing multi-model feature extraction and feature matching processing to obtain a material defect representation vector set and a defect semantic feature set; constructing a self-adaptive connection structure and a hierarchical defect map; performing multi-layer information transmission and topological relation explicit modeling through a depth map neural network to obtain a defect node depth representation set and a material defect relation matrix; domain invariant feature extraction and structural consistency constraint are carried out, material-independent defect type distribution and defect severity quantification results are obtained, defect features of all angles of the surface of the flame-retardant material can be comprehensively captured, remote interaction characteristics between defects are extracted, and the discrimination capability of defect feature representation is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular, to a method, device, equipment and storage medium for recognizing surface defects of flame retardant materials. Background Art

[0002] The performance evaluation after the combustion test of flame retardant materials mainly relies on the experience judgment of professionals. By visually observing surface defects to infer material properties, there are problems such as strong subjectivity, poor consistency, and low efficiency. Although some automated recognition technologies have been applied to surface defect detection, most of these technologies are for standardized industrial products and are difficult to effectively deal with the complex and variable surface defect characteristics generated after the combustion test of flame retardant materials.

[0003] After the combustion test, flame retardant materials will present various types of surface defects such as carbonized areas, crack networks, and bubble aggregations. There are complex topological dependence relationships between these defects, which directly affect the evaluation of the flame retardant performance of the materials. However, existing surface defect recognition technologies mainly focus on the distribution alignment of defect features and ignore the topological dependence relationships between different defect types, resulting in the recognition results being unable to accurately reflect the true performance of the materials. At the same time, due to the differences in the composition and structure of flame retardant materials, the same type of surface defect has obvious feature variability between different materials. Existing technologies lack effective cross-material domain adaptation ability and are difficult to construct a general flame retardant material performance evaluation model. Summary of the Invention

[0004] The present invention provides a method, device, equipment and storage medium for recognizing surface defects of flame retardant materials. The present invention can comprehensively capture the defect features of all angles on the surface of flame retardant materials, extract the long-range interaction characteristics between defects, and enhance the discriminant ability of defect feature representation.

[0005] In a first aspect, the present invention provides a method for recognizing surface defects of flame retardant materials, and the method for recognizing surface defects of flame retardant materials includes: Performing multi-angle image acquisition and preprocessing on the combustion test sample of the flame retardant material to obtain a standardized multi-view image data set; Performing multi-model feature extraction and feature matching processing on the standardized multi-view image data set to obtain a material defect representation vector set and a defect semantic feature set; Constructing a spatial feature double-constraint connection matrix based on the material defect representation vector set and the defect semantic feature set, and through material prior knowledge modulation and attention weight calculation, obtaining an adaptive connection structure and a hierarchical defect map; Inputting the adaptive connection structure and the hierarchical defect map into a deep graph neural network for multi-layer information transmission and explicit topological relationship modeling to obtain a defect node depth representation set and a material defect relationship matrix; Perform domain-invariant feature extraction and structural consistency constraints on the defective node depth representation set and the material defect relationship matrix to obtain a material-independent defect type distribution and a quantification result of the defect severity.

[0006] In a second aspect, the present invention provides a device for identifying surface defects of a flame-retardant material. The device for identifying surface defects of a flame-retardant material includes: An acquisition module, configured to perform multi-angle image acquisition and preprocessing on a combustion test sample of the flame-retardant material to obtain a standardized multi-view image data set; A feature extraction module, configured to perform multi-model feature extraction and feature matching processing on the standardized multi-view image data set to obtain a material defect representation vector set and a defect semantic feature set; A calculation module, configured to construct a spatial feature double-constraint connection matrix based on the material defect representation vector set and the defect semantic feature set, and obtain an adaptive connection structure and a hierarchical defect map through material prior knowledge modulation and attention weight calculation; A modeling module, configured to input the adaptive connection structure and the hierarchical defect map into a deep graph neural network for multi-layer information transmission and explicit modeling of topological relationships to obtain a defective node depth representation set and a material defect relationship matrix; An output module, configured to perform domain-invariant feature extraction and structural consistency constraints on the defective node depth representation set and the material defect relationship matrix to obtain a material-independent defect type distribution and a quantification result of the defect severity.

[0007] In a third aspect of the present invention, there is provided a computer device, including: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor invokes the instructions in the memory to cause the computer device to execute the above-mentioned method for identifying surface defects of a flame-retardant material.

[0008] In a fourth aspect of the present invention, there is provided a computer-readable storage medium, wherein instructions are stored in the computer-readable storage medium, and when the instructions are run on a computer, the computer is caused to execute the above-mentioned method for identifying surface defects of a flame-retardant material.

[0009] In the technical solution provided by the present invention, the present invention can comprehensively capture the defect characteristics of all angles on the surface of the flame retardant material, effectively solve the problem of uneven illumination on the surface of the flame retardant material through the adaptive histogram equalization and illumination normalization algorithms, and ensure the accuracy and integrity of defect feature extraction. It effectively solves the problem of insufficient labeled data in the defect recognition of flame retardant materials, and realizes high-quality representation learning with a small amount of labeled data by fusing the features of multiple visual models and the professional text descriptions of flame retardant materials. By combining the prior knowledge of the material and the attention mechanism, the connection relationship between different defect nodes is adaptively adjusted, effectively capturing the topological dependence relationship between the defects of the flame retardant material and overcoming the limitation of the traditional method that only focuses on the distribution of defect features. Through the residual jump connection and feature separation mechanism, the over-smoothing problem in the graph neural network is effectively alleviated, the remote interaction characteristics between defects are extracted, and the discriminative ability of defect feature representation is enhanced. Through domain adversarial training and structural consistency constraints, the domain shift problem between different flame retardant materials is solved, enabling the defect recognition method to have cross-material generalization ability and realizing material-independent general defect recognition. It can adaptively process multi-modal defect data, comprehensively evaluate multi-dimensional indexes such as carbonization degree, structural integrity, and thermal stability, providing a comprehensive and objective quantitative basis and material optimization direction for the performance evaluation of flame retardant materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0011] Figure 1 It is a schematic diagram of the steps of the method for identifying surface defects of a flame retardant material in an embodiment of the present invention; Figure 2 It is a schematic diagram of the structure of the device for identifying surface defects of a flame retardant material in an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0012] An embodiment of the present invention provides a method, apparatus, device, and storage medium for identifying surface defects of a flame retardant material. The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims, and above-mentioned drawings of the present invention are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the term "comprising" or "having" and any variation thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Please refer to Figure 1 , an embodiment of the method for identifying surface defects of a flame retardant material in the embodiment of the present invention includes: Step S1: Collect and preprocess multi-angle images of the flame retardant material combustion test sample to obtain a standardized multi-view image dataset; It can be understood that the execution subject of the present invention can be a device for identifying surface defects of a flame retardant material, or a terminal or a server. Specifically, it is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0014] Specifically, the flame retardant material samples after the combustion test are placed at the center of the anti-reflection turntable. The turntable has a uniform rotation function and can stably complete a rotation at a set angular velocity, so that each angle of the sample surface faces the external camera system in turn during the rotation. Five high-resolution industrial cameras are arranged around the turntable and distributed at the top, top, side, bottom and bottom positions according to a fixed spatial angle. The angles between the cameras are kept consistent to form a multi-view layout covering the sample surface. In order to eliminate the uneven illumination and shadow interference during the acquisition process at different angles, each camera is equipped with a high-brightness ring LED fill light device to ensure that the images taken at different viewing angles have stable and consistent lighting conditions. The parameters of each camera are uniformly set, including image resolution, aperture size, shutter speed and sensitivity, to ensure the consistency and comparability of the collected images. During the acquisition process, as the turntable rotates, the camera is synchronously shot at a high frequency through an external trigger mechanism to obtain a set of original multi-angle image data. The original image dataset is automatically preprocessed, and the image contrast is enhanced by the adaptive histogram equalization method, so that the local details of the material surface caused by combustion are clearer and the visual discernibility is improved. After the contrast enhancement is completed, the bilateral filtering method is used to effectively suppress the small noise points in the image, while retaining the edge contour features as much as possible to avoid the blurring and loss of detail information. After the image is denoised, image registration and geometric correction are performed to solve the perspective deformation and position offset problems caused by multi-view shooting. The images taken at different angles are uniformly adjusted to the standard reference coordinate system to obtain the corrected image. The corrected image is subjected to multi-scale segmentation processing, and each corrected image is divided into small blocks with fixed size and local overlap, so as to enhance the ability of local feature extraction and improve the recognition sensitivity of defects in the boundary area, and obtain the preliminary feature map. The preliminary feature map is subjected to illumination normalization processing. By calculating the relative ratio between the local brightness of each image area and the illumination trend of the whole image, the overall brightness of the image is adjusted to reduce the impact of uneven illumination, and a standardized multi-view image dataset is obtained.

[0015] Step S2, performing multi-model feature extraction and feature matching processing on the standardized multi-view image data set to obtain a material defect representation vector set and a defect semantic feature set; Specifically, the standardized multi-view image dataset is input into multiple pre-trained contrastive learning vision-language models, such as the vision-text multi-modal network, which have the ability to jointly model image and semantic information. Each image block is analyzed one by one through these models to extract multiple groups of preliminary image feature data. These feature vectors carry low-level visual information such as the surface topography, texture, and gray-scale distribution of the material, and also have the ability of high-level abstract expression associated with semantics. To enhance the recognition model's perception and understanding ability of the specific defect types on the surface of flame-retardant materials, relying on domain expert knowledge, a semantic description set covering multiple typical defect types is constructed. This description set contains multiple semantic terms such as carbonized texture, crack network, bubble deposition, local ablation, and holes, and is transformed into processable natural language short sentences. The text encoder in the homologous pre-trained vision-language model is used to encode the above semantic descriptions to generate a defect semantic feature set with a high-dimensional space structure. This feature set serves as a semantic anchor point in the image feature space, contains language information, and also implicitly carries the type labels and semantic boundaries of the defects, which helps to guide the orderly division of the image feature space. The similarity between the preliminary image feature data and the defect semantic feature set is calculated. By calculating the semantic similarity between the image features and the semantic anchor points, a feature similarity distribution matrix is constructed, which overall describes the matching relationship and distribution trend of each image block in the semantic space. Based on the feature similarity distribution matrix, a contrastive learning objective function is introduced in the training process. This function strengthens the intra-class compactness by minimizing the feature distance between similar image samples, and at the same time ensures the inter-class discriminability by maintaining the boundary interval between different classes, thereby continuously optimizing the distribution structure of the image features in the high-dimensional space to make it more in line with the defect classification requirements. The image features in the optimized feature space are input into the feature transformation network to perform non-linear transformation. This network adopts a multi-layer fully connected structure. Through layer-by-layer feature compression and the introduction of activation functions, it realizes the representation mapping from high-dimensional to low-dimensional, and at the same time improves the compactness and discriminability of the representation. To avoid the model overfitting to certain local features or perspective biases during the training process, a block dropout regularization mechanism is introduced into the network, that is, during the training process, some intermediate features are randomly masked with a certain probability, so that the network learns to maintain the expression stability under the condition of missing information. Through the above non-linear transformation and regularization constraints, a set of material defect representation vectors is obtained.

[0016] Step S3: Based on the set of material defect representation vectors and the defect semantic feature set, construct a spatial feature double-constraint connection matrix, and through material prior knowledge modulation and attention weight calculation, obtain an adaptive connection structure and a hierarchical defect map; Specifically, based on the distances of vectors in the material defect representation vector set in the spatial coordinates and their similarity in the feature space, while considering physical location proximity and semantic feature proximity, a dual-constrained connection matrix of spatial features is comprehensively constructed. Each element in this matrix reflects the possibility and strength of establishing a connection between two defect regions in both the spatial and semantic dimensions. The material defect representation vector set is input into the linear transformation module. By calculating the query matrix and the key-value matrix for each node vector respectively, a defect self-attention matrix is generated. This matrix is used to characterize the degree of mutual attention between nodes in the high-dimensional semantic space, that is, the attention weight of a certain defect region to other defect regions in the comprehensive analysis. At the same time, based on the semantic anchors constructed from the defect semantic feature set, from the perspective of materials science, a material prior knowledge matrix is constructed based on expert knowledge and empirical rules. Each element in this matrix represents the association strength between two types of defects in the material physical mechanism. For example, the association degree between the carbonization region and the crack is much higher than that between the carbonization region and surface contamination, so the corresponding element value is larger. The element-wise product operation is performed on the defect self-attention matrix and the material prior knowledge matrix, enabling the model to embed prior knowledge in the context of materials expertise while paying attention to feature similarity, and constructing a modulated connection weight matrix. To enhance its non-linear modeling ability, non-linear activation processing is performed on the product result, such as normalizing and compressing it through a gating mechanism or an activation function, so that the formed connection modulation weight matrix not only has perception ability but also has generalization and selectivity. The modulated weight matrix and the dual-constrained connection matrix of spatial features are again subjected to element-wise product operation to obtain an adaptive connection structure. The material defect representation vector set is clustered according to feature similarity, and the nodes are hierarchically divided according to the clustering results. On this basis, the in-layer graph structure and the inter-layer connection strength are calculated respectively, and a hierarchical defect map with clear semantic boundaries and topological levels is constructed.

[0017] In this embodiment, taking the material defect representation vector set as the input, the relationships between each pair of representation vectors are calculated in full, and the Euclidean distance is used as the metric standard. The geometric distances between any two vectors in the feature space are measured in sequence to generate a feature similarity matrix, which reflects the degree of proximity of all defect regions in the high-dimensional feature space. Based on the feature similarity matrix, spectral clustering analysis is performed on the material defect representation vector set. By performing Laplacian eigen-decomposition on the similarity matrix, the complex sample space is mapped to a new low-dimensional embedding space, in which samples with similar features can be effectively aggregated. Considering the complexity and diversity of the characterization of flame-retardant material defects, the number of clusters is set to four, and four defect feature clusters are divided, and different feature expression dimensions are assigned to each cluster, corresponding to 64 dimensions, 128 dimensions, 256 dimensions, and 512 dimensions respectively, thereby constructing a multi-scale feature cluster system with a gradually progressive expression ability. For the nodes within each feature cluster, that is, the defect representation vectors belonging to the same layer, a local adjacency relationship network is constructed through a neighborhood search strategy, and four intra-layer connection structure diagrams are generated on the premise of maintaining the original structural information of the nodes. These structure diagrams respectively depict the local topological structures of each type of defect in the same-layer feature space. At the same time, to capture the potential global dependence relationships between cross-layer nodes, the inter-layer connection weights between different feature clusters are calculated, and the connection strength is determined by comparing the feature correlations between cross-layer nodes. For example, pairwise feature similarity matching is performed between the nodes in the upper layer and the nodes in the lower layer, and the matching scores are aggregated and normalized to serve as the inter-layer connection weights, and an inter-layer connection matrix is constructed, which reflects the structural mapping relationship and information transmission channels between each layer. The four intra-layer connection structure diagrams and the obtained inter-layer connection matrix are integrated to construct an initial hierarchical graph structure, which has the ability of cross-layer connection in the vertical direction and the same-layer feature propagation mechanism in the horizontal direction, and can capture both the local details and the global structure of the defects. The self-attention mechanism is used to optimize the inter-layer information flow of the initial hierarchical graph structure. During the propagation process of the graph neural network, an attention weight is assigned to each node, and this weight is dynamically adjusted according to the semantic similarity and structural importance between the node and its upper and lower layer neighbor nodes, so that a larger information channel weight is obtained between nodes with strong dependence relationships, while redundant or noisy connections are automatically compressed or ignored. Through multiple rounds of iterative calculation and feature update of the self-attention mechanism, the semantic alignment between layers is gradually enhanced, and a hierarchical defect atlas with a clear structural distribution, distinct semantic levels, and reasonable information flow is constructed.

[0018] Step S4: Input the adaptive connection structure and the hierarchical defect atlas into a deep graph neural network for multi-layer information transmission and explicit modeling of topological relationships, and obtain a set of deep representations of defect nodes and a material defect relationship matrix; Specifically, the adaptive connection structure and the hierarchical defect map are input into the depth node graph representation learning module of the deep graph neural network for node feature extraction. The multi-dimensional attribute information of the nodes propagates and fuses in the graph structure to form initial multi-layer node features across levels. Using the topological weight relationship provided by the adaptive connection structure, aggregate calculations are performed on the neighbor set of each node, and a bilinear attention mechanism is introduced for neighborhood information weight reconstruction. This mechanism calculates the combined feature expression between the target node and its adjacent nodes, non-linearly processes the attention scores using the LeakyReLU activation function, and further multiplies them by the connection strength weights to obtain neighborhood information transfer weights with semantic selection capabilities. The initial multi-layer node features are input into the feature transfer network in the deep graph neural network, and the node information of each layer is fused and reconstructed through a multi-layer convolution propagation mechanism, enabling the features of each layer to retain the local characteristics of this layer and also receive global semantic enhancement from the upper layer. After multiple rounds of propagation and fusion, a set of stable and structurally consistent multi-layer fusion features is obtained. A feature decoupling network is introduced to perform projection processing on the multi-layer fusion features. This network guides the gradient to propagate along a specific semantic direction, distinguishes shared features from class-specific features, and effectively compresses redundant feature dimensions through an orthogonal projection mechanism in the gradient space, outputting a set of defect node depth representations with high discriminability. The set of defect node depth representations is input into the graph projection network for defect relationship modeling. The network maps the node representations to a predefined semantic space, calculates the relationship strength between any two nodes, and forms an initial matrix of defect type relationship strengths, which reflects the degree of structural association between various types of defects learned by the system. Calculate the Frobenius norm difference between the initial matrix of defect type relationship strengths and the ideal relationship matrix constructed based on flame retardant material science knowledge to quantify the structural deviation between the current relationship modeling result and the actual understanding of the material. By optimizing the gradient descent of this difference index during the training process, the network continuously adjusts its parameters to minimize the distance between the two, obtaining a material defect relationship matrix that conforms to both data characteristics and physical mechanisms.

[0019] Step S5: Perform domain-invariant feature extraction and structural consistency constraint on the set of defect node depth representations and the material defect relationship matrix to obtain a material-independent defect type distribution and a quantified result of the defect severity.

[0020] Specifically, the depth representation set of defective nodes is input into the gradient reversal layer to construct a domain classifier for training. By introducing the gradient reversal layer, this classifier realizes adversarial learning of material domain information in the feature space. That is, during training, on the one hand, it encourages the main network to retain information beneficial to defect recognition, and on the other hand, it forces the domain classifier to be difficult to accurately judge the material type to which the feature belongs. As a result, the finally output features are made to eliminate the differences between different material domains as much as possible in terms of distribution, making them have good generality and transfer ability. In this process, by minimizing the cross-entropy loss of the domain classifier, the extraction of domain-invariant feature representations is achieved. At the same time, to ensure that the topological characteristics of the defective graph structures between different materials remain consistent, a structure consistency constraint mechanism is introduced. Based on the depth representation set of defective nodes and the material defect relationship matrix, the Laplacian matrix and connection matrix of the graph are constructed on the source domain and the target domain respectively, so as to extract the respective graph topological structure information. The norm difference of the corresponding graph matrices between the source domain and the target domain is calculated, and by minimizing the difference metric, the network is constrained to maintain structure consistency during the learning process, preventing the graph structure from degrading or the semantic transfer from being unstable due to differences in material physical properties. This structure consistency representation serves as an intermediate supervision signal to guide the model to learn a stable structure encoding method across material domains, thereby enhancing the model's topological perception ability of defects in different materials. According to the attribute information of different material samples, including density, thickness, hierarchical structure, etc., the attribute similarity between material domains is calculated, and this is used as the inter-domain similarity weight, which acts on the norm calculation process between the material defect relationship matrices, making the structure alignment strength higher between similar materials, while dissimilar materials retain their differences, forming a controllable multi-domain structure adversarial optimization objective. Through the minimization training strategy of this multi-domain adversarial loss, the network can adapt to the feature transfer process between different materials and perform unified modeling on multi-material data. The domain-invariant feature representation is input into the graph convolutional classifier, and the type of defective nodes is predicted through non-linear mapping. While using the graph structure to enhance the feature discrimination ability, this classifier has the ability to embed the input features into the standard defect semantic space, so as to output the defect type distribution independent of the specific material type, ensuring stable classification performance in scenarios with complex sample sources. To achieve a quantitative assessment of the defect severity, the domain-invariant feature representation and the structure consistency representation are jointly input into a specially designed regression network. This network models the association between local features and the global structure and outputs the severity score corresponding to each defective node. This score uses a continuous numerical form to represent the impact degree of the defect and has high interpretability and evaluation reference value.

[0021] In the embodiments of the present invention, the present invention can comprehensively capture the defect characteristics of the surface of the flame retardant material at various angles, effectively solve the problem of uneven illumination on the surface of the flame retardant material through the adaptive histogram equalization and illumination normalization algorithms, and ensure the accuracy and integrity of defect feature extraction. It effectively solves the problem of insufficient labeled data in the defect recognition of flame retardant materials, and realizes high-quality representation learning with a small amount of labeled data by fusing the features of multiple visual models and the professional text descriptions of flame retardant materials. By combining the prior knowledge of the material and the attention mechanism, the connection relationship between different defect nodes is adaptively adjusted, effectively capturing the topological dependence relationship between the defects of the flame retardant material, and overcoming the limitation of the traditional method that only focuses on the distribution of defect features. Through the residual jump connection and feature separation mechanism, the over-smoothing problem in the graph neural network is effectively alleviated, the long-range interaction characteristics between defects are extracted, and the discriminative ability of defect feature representation is enhanced. Through domain adversarial training and structural consistency constraints, the domain shift problem between different flame retardant materials is solved, enabling the defect recognition method to have cross-material generalization ability and realizing material-independent general defect recognition. It can adaptively process multi-modal defect data, comprehensively evaluate multi-dimensional indicators such as carbonization degree, structural integrity, and thermal stability, providing a comprehensive and objective quantitative basis and material optimization direction for the performance evaluation of flame retardant materials.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Place the combustion test sample of the flame retardant material at the center of the anti-reflection turntable, and perform multi-angle image acquisition on the combustion test sample of the flame retardant material through five high-resolution industrial cameras distributed at fixed angles to obtain the original multi-angle image data; Perform adaptive histogram equalization processing on the original multi-angle image data to obtain a contrast-enhanced image, and perform bilateral filtering processing on the contrast-enhanced image to obtain a noise-reduced image; Perform image registration and geometric correction processing on the noise-reduced image to obtain a corrected image, and perform multi-scale segmentation processing on the corrected image to obtain a preliminary feature map; Perform illumination normalization processing on the preliminary feature map to obtain a standardized multi-view image dataset.

[0023] Specifically, place the combustion test sample of the flame retardant material at the center of the anti-reflection turntable. Utilize the uniform rotation characteristic of the turntable to ensure that each surface area of the sample faces the imaging device in a stable state during the image acquisition process. The anti-reflection turntable enhances the consistency and clarity of image acquisition by reducing background reflection interference and avoiding the influence of non-uniform external light incidence, thus ensuring homogeneous performance of the images at various angles. Five high-resolution industrial cameras are evenly arranged around the turntable center. Each camera is precisely positioned at different spatial angles through a fixed bracket, including 0 degrees directly above, 30 degrees above the side, 90 degrees on the side, 150 degrees below the side, and 180 degrees at the bottom, forming a three-dimensional shooting network that comprehensively covers the upper surface, edges, and bottom of the sample. Each camera is equipped with a high-brightness, stable-power ring-shaped LED lighting device, and the lighting intensity is controlled at a constant 1500 lumens. By controlling the stability of the ambient lighting and the camera light source, the problems of image contrast fluctuation and reflection interference caused by changes in external lighting conditions are effectively solved. All cameras adopt the same parameter configuration, including the resolution set to 3840×2160 pixels, the aperture value fixed at F2.8, the shutter speed uniformly set to 1 / 100 second, the ISO sensitivity controlled within 400, and the image saving format selected as the lossless RAW format to retain complete original information. As the turntable rotates uniformly at a speed of 10 degrees per second, each camera captures 3 frames of images per second according to the synchronous trigger signal. The entire process lasts for about 12 seconds, obtaining approximately 108 evenly distributed and angle-complete original multi-angle image datasets. Perform adaptive histogram equalization processing on the original multi-angle image data to enhance the local contrast of the images, so that microscopic defect features on the surface of the burned material, such as cracks, carbonization marks, and blistering areas, can be effectively enhanced under different lighting backgrounds. The adaptive method dynamically adjusts based on the local gray distribution of the image region, and carefully compensates for the areas with insufficient brightness or excessive concentration in the image, thus ensuring the unity and rich details of the overall visual effect of the image. After the contrast enhancement is completed, perform bilateral filtering operation on the image. This filtering technology performs weighted averaging on the image pixels by combining two factors of spatial distance and pixel similarity, effectively suppressing random noise. Especially in areas with strong texture or obvious edges, it can reduce high-frequency noise while retaining clear boundaries, avoiding damage to the tiny defect structure by conventional smoothing filtering, thereby improving the usability of the image quality. After the image denoising process, to ensure the consistency of multi-view images in the spatial coordinate system, perform registration and geometric correction processing on all images. By extracting key points in the images and calculating the perspective transformation matrix, the reconstruction of images at different angles in a unified reference coordinate system is realized. Especially during the movement of the turntable, there are perspective offsets and projection distortions in the images captured by each camera. Through geometric correction, deformation problems such as bending, tilting, and stretching of the sample edges are eliminated, ensuring the precise alignment of the same defect area in images from different perspectives, thus providing a position reference basis for subsequent tile analysis and cross-image fusion.After the registration and calibration are completed, the image enters the multi-scale image segmentation stage. The large image is divided into small image patches suitable for sliding window processing. Each image is divided into multiple 128×128 pixel blocks, and a 25% overlapping area is set between adjacent image patches to enhance the expression ability of local continuous features. The generated preliminary feature map is input into the illumination normalization module for processing to eliminate the uneven brightness phenomenon caused by shooting at different angles or the reflection conditions of different material surfaces, and to improve the illumination consistency of the image in the global and local ranges. Illumination normalization is achieved by constructing a set of normalization functions based on the ratio of local intensity to global trend, using Gaussian filtering to extract the global illumination distribution of the image, and then dividing it by the local pixel values to complete the illumination correction, so that all regions of the image, regardless of their original illumination intensity, are finally adjusted to a unified visual brightness standard, obtaining a standardized multi-view image dataset.

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Input the standardized multi-view image dataset into multiple pre-trained contrastive learning vision-language models for image block analysis to obtain preliminary image feature data; Construct a semantic description set of flame retardant material defects according to the professional knowledge of flame retardant material defects, and perform text feature encoding on the semantic description set of flame retardant material defects to obtain a defect semantic feature set; Calculate the similarity between the preliminary image feature data and the defect semantic feature set to obtain a feature similarity distribution matrix; Construct a contrastive learning objective function based on the feature similarity distribution matrix, and obtain an optimized feature space by minimizing the intra-class sample distance and maintaining the inter-class boundary interval; Input the image features in the optimized feature space into the feature transformation network to perform non-linear transformation and block dropout regularization processing to obtain a set of material defect representation vectors.

[0025] Specifically, the standardized multi-view image dataset is input into multiple pre-trained contrastive learning vision-language models for feature analysis at the image block level. The pre-trained models are trained based on large-scale text-image alignment corpora and possess strong cross-modal embedding capabilities and feature abstraction and expression capabilities. Each image sub-block obtained by multi-scale division is respectively input into a model composed of a vision Transformer structure. Under the combined action of different model structures, initial image feature vectors with multiple expression dimensions and different modeling granularities are extracted. These features include low-level visual information such as material surface texture, edge morphology, and local brightness, as well as deep structural features of the residues after material combustion, forming an initial image feature dataset with unified dimensions and rich semantics. Since this process is based on parallel extraction by multiple models, each image block will be separately encoded into a group of high-dimensional feature vectors. A semantic description set for flame-retardant material defects is constructed according to the professional knowledge of flame-retardant material defects. This semantic description set consists of multiple term short sentences, covering typical surface defect types after combustion such as "carbonization cracks", "surface bubbling", "tiny holes", "edge delamination", "structural ablation", and "cracked texture". Each term description is regarded as a semantic anchor point to guide the aggregation of image features in an interpretable direction. The semantic description set is input into a text encoder homologous to the image model for feature encoding, so that the language description is embedded into the same feature space in the form of text vectors, generating a defect semantic feature set. Since the vision-language structure used in the encoding model has cross-modal alignment capabilities, the generated text vectors have the ability to maintain semantic connotations and align with visual features, serving as category reference points in the image feature space. The similarity between the initial image feature data and the defect semantic feature set is calculated. The cosine similarity or dot product in the vector space is used as the matching criterion to pair and match each image block feature with all semantic anchor points, forming an image-semantic feature similarity distribution matrix. This matrix records the matching strength between image samples and all semantic labels and reflects the possible distribution trend of samples in the semantic space. Based on the feature similarity distribution matrix, a contrastive learning objective function for the image recognition task is constructed, taking the similarity between image features and semantic features as the core metric. By designing an optimization mechanism for intra-class aggregation and inter-class separation, the reconstruction and enhancement of the feature space structure are realized. During the optimization process, the system continuously minimizes the feature distance between image samples under the same semantic category, while setting a minimum category interval to maintain the clarity of the boundaries between different categories, ensuring the aggregation of similar samples and the separation of different samples, thereby improving the discriminative ability and semantic aggregation ability of image features, making the feature embedding space tend to be stable under dual visual and semantic constraints, and enhancing its sensitivity to defect types to obtain an optimized feature space. The image features in the optimized feature space are input into a feature transformation network to perform non-linear transformation and block dropout regularization processing.The feature transformation network consists of multiple non - linear transformation structures, including a series of linear mapping layers, activation functions, and normalization modules. It maps the initial high - dimensional image feature vectors into a set of more compact and task - relevant low - dimensional representations, reducing redundancy and computational burden while maintaining expressiveness. During feature transformation, a regularization strategy of block dropout is introduced in the training stage, that is, randomly masking some intermediate layer nodes or sub - block features during training. By artificially creating a feature - missing situation, the model is forced to learn a more robust and redundancy - controllable global expression structure, preventing it from overfitting to local features or noise information. A set of deep vectors with semantic representation ability, structural consistency, and cross - material adaptability is obtained, namely the material defect representation vector set.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Construct a spatial feature double - constraint connection matrix according to the spatial positions and feature distances of the vectors in the material defect representation vector set; Input the material defect representation vector set into a linear transformation module to calculate the query matrix and key - value matrix, obtaining a defect self - attention matrix; Construct a material prior knowledge matrix based on the defect semantic feature set; Perform an element - wise product operation on the defect self - attention matrix and the material prior knowledge matrix and perform non - linear activation processing to obtain a connection modulation weight matrix; Perform an element - wise product operation on the spatial feature double - constraint connection matrix and the connection modulation weight matrix to obtain an adaptive connection structure; Cluster the material defect representation vector set by feature similarity in layers and calculate the inter - layer connection strength to obtain a hierarchical defect map.

[0027] Specifically, a spatial feature dual-constraint connection matrix is constructed based on the spatial positions and characteristic distances of the vectors in the material defect representation vector set. The Euclidean space distance matrix between all defect nodes is calculated to measure the proximity of each node in the physical image space; at the same time, the characteristic distances between the defect representation vectors are evaluated to judge the similarity strength in the deep semantic feature space. These two are used as independent indicators and are jointly integrated into the spatial feature dual-constraint connection matrix after unified normalization. Each element of this matrix represents the connection probability between two nodes under the combined action of spatial proximity and semantic similarity, forming the basis for the initial edge weight setting of the graph structure and avoiding the connection distortion problem caused by the traditional single distance measurement method ignoring semantic relationships or spatial continuity. The material defect representation vector set is input into the linear transformation module for calculating the query matrix and the key-value matrix. The vectors of each defect node are linearly mapped to obtain a query vector representing its "request feature" and key vectors and value vectors representing its "received features", and based on this, the self-attention score matrix between each node is calculated through inner product. The defect self-attention matrix reflects the ability of nodes to actively perceive each other in the feature space, that is, how the attention degree of each node to all other nodes is distributed, thus establishing an asymmetric, information-dependent graph connection weight expression method. Through the self-attention mechanism, the model does not rely on fixed spatial or semantic rules for connection, but adjusts the connection strength and information transmission channels between it and other nodes in real time according to the dynamic changes of each node's features, thereby enhancing the adaptability and expression ability of the graph structure under complex defect morphologies. A material prior knowledge matrix is constructed based on the defect semantic feature set. The design of this matrix comes from professional cognitions such as the defect formation causes, symbiotic relationships, and structural evolution mechanisms in the field of flame-retardant materials. For example, carbonized areas are accompanied by crack propagation, and bubble aggregation is likely to cause local delamination, etc. These empirical rules are quantitatively encoded as the association weights between nodes and act together with the self-attention results in matrix form. The defect self-attention matrix and the material prior knowledge matrix are subjected to element-wise multiplication operation, that is, the attention scores and the expert prior weights at the same position are fused, so that the connection weights not only retain the dynamic perception structure learned by the network itself but also embed the guiding constraints of material physical laws. A non-linear activation operation is performed on the fused matrix. By introducing activation functions such as Sigmoid or ReLU, the original continuous weights are mapped to the interval range, making the connection strength exhibit saturation and selectivity characteristics, forming a connection modulation weight matrix. The initially constructed spatial feature dual-constraint connection matrix and the connection modulation weight matrix are subjected to element-wise multiplication operation again to obtain an adaptive connection structure that integrates spatial, feature, self-attention, and material prior information. The material defect representation vector set is hierarchically clustered according to feature similarity and the inter-layer connection strength is calculated.Using clustering algorithms, such as spectral clustering or hierarchical clustering, classify the node vectors according to their distribution density and mutual distance in the semantic feature space, and divide them into several clusters with highly similar structures. Each cluster is regarded as a structural layer in the defect map, representing a relatively homogeneous defect group. For example, the carbonized areas are classified into one level, while the edge cracks or blister structures are classified into other different levels. After completing the clustering, calculate the connection strength between each two levels. This connection strength is determined by the feature correlation between the cross-layer nodes. By calculating the cosine similarity or Euclidean distance difference of the node feature means between different clusters, and combining the edge weight data in the adaptive connection structure, perform aggregation statistics to obtain the inter-layer connection matrix. Integrate the intra-layer connection structure and the inter-layer connection matrix to construct a hierarchical defect map.

[0028] In a specific embodiment, the process of performing the steps of hierarchically clustering the material defect representation vector set according to feature similarity and calculating the inter-layer connection strength to obtain a hierarchical defect map may specifically include the following steps: Calculate the Euclidean distance for each vector in the material defect representation vector set to obtain a feature similarity matrix; Based on the feature similarity matrix, perform spectral clustering analysis on the material defect representation vector set to obtain four layers of defect feature clusters, and the feature dimensions of the four layers of defect feature clusters are 64, 128, 256, and 512 respectively; Construct an adjacency relationship for the nodes within each layer of defect feature clusters to obtain four intra-layer connection structure diagrams; Calculate the connection weights between the four layers of defect feature clusters according to the feature correlation between the cross-layer nodes to obtain an inter-layer connection matrix; Integrate the intra-layer connection structure diagram and the inter-layer connection matrix to obtain an initial hierarchical graph structure; Use the self-attention mechanism to optimize the inter-layer information flow of the initial hierarchical graph structure to obtain a hierarchical defect map.

[0029] Specifically, Euclidean distance calculations are performed on each vector in the material defect representation vector set to measure the degree of their distribution differences in the high-dimensional semantic space. Each vector represents the compressed representation of an image block in the visual semantic space, containing comprehensive features such as local texture, structural boundaries, burn marks, and potential defect types. The calculation of Euclidean distance is regarded as a quantitative description of the difference degree of defect characteristics of different image blocks. By performing a full calculation among all vectors, a symmetric feature similarity matrix is obtained, where each element in the matrix represents the similarity degree between two defect samples in the feature space. Spectral clustering analysis is performed on the material defect representation vector set based on the feature similarity matrix. Spectral clustering maps the high-dimensional non-linear distribution to a linearly separable new space by solving the Laplacian matrix and performing eigen-decomposition operations on the similarity matrix. In this space, by determining the aggregation degree of the projection vectors of the nodes, multiple sub-cluster groups with compact structures and high similarities are divided. To adapt to the complexity of defect distribution and the expression requirements of multi-scale feature structures, the system pre-sets the clustering level to four layers, dividing four defect feature clusters, which respectively correspond to the structural expression layers of four different semantic dimensions. The feature dimensions of these layers are 64, 128, 256, and 512 in sequence, reflecting the gradual enhancement of representation from low-level visual patterns to high-level semantic structures. After clustering, the system constructs the local connection structure for the nodes within each layer, that is, within the same defect cluster, connection edges are set and weights are assigned according to the feature similarity and spatial adjacency relationship between nodes, forming four relatively independent intra-layer connection graph structures. Each graph structure reflects the close connection between defects of this type in space or semantics, constituting the basic structural unit of the hierarchical graph. The connection weights between the four-layer defect feature clusters are calculated according to the feature correlation between cross-layer nodes, and an inter-layer connection matrix is constructed. Analyze the feature correlation between cross-layer nodes. On the one hand, the initial inter-layer relationship distribution is obtained by calculating the similarity between the central feature vectors of nodes in different layers. On the other hand, consider the defect co-occurrence mechanism existing between cross-layers. For example, some high-temperature carbonization regions are located above the crack extension area, or the bubble aggregation layer appears between the delamination and ablation boundaries. These relationships are indirectly reflected through the embedded representation between nodes. The system synthesizes the above information, calculates the connection weights of all possible cross-layer combinations between the four layers, and expresses this structural relationship in matrix form, forming an inter-layer connection matrix. Each element of this matrix represents the feature correlation degree and potential information transfer ability between two different layers. Integrate the intra-layer connection structure diagram and the inter-layer connection matrix. Combine the four intra-layer graphs and the inter-layer connection matrix to form an initial hierarchical graph structure. This graph structure has the semantic aggregation characteristics between nodes in the horizontal dimension and reflects the topological dependence relationship between layers in the vertical dimension, forming a composite graph model of "intra-layer detail representation + inter-layer structure induction".Considering that the initially constructed hierarchical structure still belongs to a static connection structure and has not dynamically optimized the information flow path, in order to improve the efficiency of information propagation between layers and avoid redundant conduction of low-correlation paths, a self-attention mechanism is introduced to optimize the information flow regulation of the entire hierarchical graph structure. During the propagation process of the graph neural network, a multi-head self-attention mechanism is adopted to dynamically calculate the information importance weights between different layers. Each node not only interacts with its neighbors in the same layer but also establishes a weighted information transmission channel with relevant nodes in the upper and lower layers according to the self-attention mechanism. The calculation of the attention weights is based on the feature vector correlation between nodes, the connection structure strength, and their distance distribution in the clustering space. By fusing and modeling these dimensions, different transmission weights are assigned to the information channels between each layer, and a non-linear compression is performed using an activation function to ensure that important channels obtain sufficient transmission capacity while redundant paths are effectively suppressed. In the multi-round iterative information propagation, the self-attention mechanism continuously optimizes the inter-layer connection structure and node representation, forming a hierarchical defect map with semantic aggregation, structural clarity, and multi-scale perception capabilities.

[0030] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Input the adaptive connection structure and the hierarchical defect map into the deep node graph representation learning module of the deep graph neural network for node feature extraction to obtain initial multi-layer node features; Calculate the attention scores for the neighborhood of each node based on the adaptive connection structure, and multiply them by the connection weights through a bilinear attention mechanism activated by LeakyReLU to obtain the neighborhood information transmission weights; Input the initial multi-layer node features into the feature transfer network of the deep graph neural network for feature fusion to obtain multi-layer fusion features; Input the multi-layer fusion features into the feature decoupling network for gradient projection to obtain a set of deep representations of defect nodes; Input the set of deep representations of defect nodes into the graph projection network for relationship analysis to obtain an initial matrix of defect type relationship strengths; Calculate the Frobenius norm difference between the initial matrix of defect type relationship strengths and the ideal relationship matrix constructed based on the knowledge of flame retardant materials science, and optimize it through gradient descent iteration to obtain the material defect relationship matrix.

[0031] Specifically, the adaptive connection structure and the hierarchical defect map are input into the depth node graph representation learning module of the deep graph neural network for node feature extraction. By combining structural connection and node feature representation, the initial multi-layer feature encoding of each defect node in the graph is completed. At this stage, the network not only considers the feature information of each node itself, but also integrates the relevant information of its adjacent nodes in the local semantic space. Based on the combination of topological structure and feature distribution, the initial node feature representation of each layer is output, and the graph node expression structure at different semantic scales is established. Calculate the attention score for the neighborhood of each node based on the adaptive connection structure. During the feature propagation process, the network constructs a set of attention functions based on the bilinear form for each node and its adjacent node pair. Through this function, the features between the target node and the neighbor node are calculated by cross-combining, and the potential semantic dependence strength is extracted. At the same time, to introduce a non-linear response mechanism, a LeakyReLU activation operation is connected to the output end of the attention function to solve the problem of gradient disappearance caused by overly sparse eigenvalue distribution and enhance the network's response ability to weak feature paths. Through this attention mechanism, the information perception intensity of each target node for its neighborhood nodes is obtained. After element-wise multiplication with the connection weights in the adaptive connection structure, the neighborhood information transfer weight matrix is obtained. Input the initial multi-layer node features and the calculated neighborhood information transfer weights into the feature transfer network of the deep graph neural network. In this network, each layer is controlled by a learnable weight matrix for the transformation manner of node features, and combined with the attention weights calculated in the adjacent graph, the weighted aggregation of adjacent node features is realized. Through multi-layer transfer and fusion operations, the system gradually superimposes, fuses, and updates the local features of the nodes with the information in their surrounding structures, enabling the nodes to have stronger global structure perception ability and category feature aggregation ability while maintaining their own semantic expressions. This fusion process adopts a residual connection and normalization mechanism to prevent over-smoothing or gradient disappearance in the deep network, and realizes parallel perception of different semantic patterns through a multi-head structure, outputting a set of multi-layer fusion feature representations with both detail resolution and structure alignment capabilities. To improve the representation ability of node features in category discrimination and structure modeling tasks, a feature decoupling network is introduced after feature fusion, and a gradient projection mechanism is introduced in this network to achieve structural separation and orthogonal expression in the feature space. The core task of the decoupling network is to decompose the fused features of each node into a shared part and a category-specific part. The shared part captures the common geometric or texture structures across categories, while the category-specific part reflects the unique semantic features of each type of defect. Through the orthogonal projection operation in the gradient space, the network strongly constrains these two types of feature dimensions to prevent cross-interference or representation redundancy in the expression space, obtaining a set of highly distinguishable and clearly expressed defect node depth representation vector sets. Input the defect node depth representation set into the graph projection network to complete the relationship modeling and graph reasoning among defects.The graph projection network, as the core part of the graph structure modeling module, is used to map the semantic representation of nodes into a structure space specifically for representing the relationships between defect types. In this space, the network calculates an initial matrix of defect type relationship strengths through mutual projection and similarity comparison between pairs of node feature pairs. Each element of this matrix reflects the strength of the internal connection between two defect types in the graph structure semantic space. The larger the matrix value, the more likely it indicates that the two types of defects have a co-evolution or physical adjacency relationship during the actual material damage or combustion process; otherwise, they are relatively independent or mutually exclusive. Calculate the Frobenius norm difference between the initial matrix of defect type relationship strengths and the ideal relationship matrix constructed based on the knowledge of flame-retardant materials science. The ideal relationship matrix is compiled by materials experts based on years of experimental data and theoretical analysis results, reflecting the objective connections between material defect types in terms of physical causes, structural dependencies, spatial evolution, etc. To achieve the integration of data-driven modeling and knowledge-driven modeling, by calculating the Frobenius norm difference between the initial relationship matrix and the ideal relationship matrix, obtain the structural deviation degree between the current modeling result and the actual understanding of the material, and use this deviation degree as a loss function to be added to the training objective, and perform iterative optimization through multiple rounds of gradient descent. During each optimization process, the network will automatically adjust the graph projection parameters to gradually approximate the initial relationship matrix to the ideal matrix, and obtain the material defect relationship matrix.

[0032] In a specific embodiment, the process of executing step S5 may specifically include the following steps: Input the deep representation set of defect nodes into the gradient reversal layer to construct a domain classifier and perform minimization of the domain classification cross-entropy loss calculation to obtain domain-invariant feature representations; Calculate the Laplacian matrix and connection matrix of the source domain and the target domain based on the deep representation set of defect nodes and the material defect relationship matrix, and obtain the structural consistency constraint representation by minimizing the matrix norm difference; Calculate the inter-domain similarity weight based on the attribute similarity of different flame-retardant material types, and perform weighted norm calculation on the material defect relationship matrix according to the inter-domain similarity weight to obtain the multi-domain adversarial loss value; Input the domain-invariant feature representations into the graph convolutional classifier for non-linear mapping to obtain the material-independent defect type distribution; Input the domain-invariant feature representations and the structural consistency constraint representations into the regression network for severity score calculation to obtain the quantification result of the defect severity.

[0033] Specifically, the depth representation set of defective nodes is input into the gradient reversal layer to construct a domain classifier and perform the calculation of minimizing the domain classification cross-entropy loss. The gradient reversal layer serves as a bridge connecting the main task network and the auxiliary domain discriminator. During backpropagation, it reverses the gradient sign, forcing the backbone network to learn features that are not easily recognized by the domain classifier during training, thereby effectively erasing the statistical distribution differences between different material sources. The depth representation vector of the defective nodes is used as the input and passed into the domain classifier through the gradient reversal layer. This classifier calculates the standard cross-entropy loss function with the material domain label as the supervision signal, maximizes the feature confusion by minimizing the domain classification accuracy, and obtains a set of domain-invariant feature representations that do not depend on the material source. At the same time, considering the differences in the graph structures from different material sources at the topological level, a structure consistency constraint mechanism is introduced as an auxiliary optimization objective to enhance the consistent expression of cross-material graph structures. This mechanism analyzes the graph structure performance of the depth representation set of defective nodes and the material defect relationship matrix in the source domain and the target domain, calculates the graph Laplacian matrix and the connection matrix of the two domains respectively. The graph Laplacian matrix reflects the information flow ability between nodes, while the connection matrix directly records the graph edge weights or connection relationships between nodes. By performing element-wise difference measurement on these two types of matrices, the deviation degree of the source domain and the target domain in the graph topology structure is obtained. Further, the Frobenius norm is used as the structure difference loss function, and this structure deviation is introduced into the overall training objective as a regular term. The network parameters are reversely optimized through gradient descent, prompting the model to gradually adjust the node representation and connection method during the learning process, thereby achieving graph structure alignment between different material domains at the structural level and obtaining a set of structure consistency constraint representations with consistent graph topology distribution characteristics. To enhance the adversarial learning ability of the system among multi-material domains and suppress the impact of extreme material property differences on the recognition effect, a multi-domain adversarial loss mechanism is designed. This mechanism constructs an inter-domain similarity model based on the physical and process properties of flame-retardant materials. By evaluating the property similarity between different material types, such as the Euclidean distance or statistical distribution overlap degree between parameters such as density, thickness, combustion temperature, and structural hierarchy, the similarity weight matrix between each material domain is calculated. This similarity weight reflects the physical proximity at the material property level and can also be used to adjust the relationship constraint strength between different domains. Using the similarity weight as a coefficient, the weighted norm difference calculation is performed on all material defect relationship matrices, that is, a multi-domain loss function is constructed. Its objective is to limit the excessive deviation of the structural expression while maintaining the reasonable difference of the domain structure, and improve the domain adaptation ability of the overall model and the robustness of multi-domain learning.During the training process, the multi-domain adversarial loss, together with the domain classification loss, structural consistency loss, etc., serves as an optimization objective to participate in backpropagation. Under the training strategy of multi-loss cooperation, the system continuously converges to a parameter solution space that takes into account both feature discriminability and structural stability, ensuring that the model can still stably output accurate defect types and severity levels when facing input images of multiple material types. After completing the above-mentioned feature space domain adaptation and structural space alignment, the domain-invariant feature representation is used as input to the graph convolutional classifier for defect type prediction. This classifier adopts a standard graph convolutional neural network structure, combines the adjacency relationship information in the graph with the features of the nodes themselves, and learns the joint feature representation between the nodes and their neighbors during multiple rounds of graph convolutional propagation and non-linear activation processes. Through hierarchical feature aggregation and spatial perception mechanisms, the node features are mapped into a preset defect type space. The output end uses a Softmax classifier to map each node to a specific defect type category, constructing a material-independent defect type distribution. Since the input features have been processed by domain adversarial training, the classification results are no longer affected by material sources, image perspectives, or surface interference factors, and have good stability and generalization ability, being used for defect recognition tasks of any material samples and supporting subsequent recognition evaluation and material optimization links. After the defect type recognition is completed, the domain-invariant feature representation and the structural consistency constraint representation are jointly input into the regression network to calculate the severity score of the defect. The structure of this regression network adopts a multi-layer perceptron form. At the front end, the two types of input features are spliced and fused. There are multiple fully connected layers and activation functions inside to capture the non-linear interaction relationships between different features. At the same time, techniques such as Dropout and BatchNorm are used to prevent the model from overfitting. The network output is a continuous value between 0 and 1, representing the severity score of the defect of each node. The higher the value, the greater the potential threat of the defect to the material structure performance, and the lower the value, the lighter the damage in this area and the higher the controllability. The regression model is trained through a supervised learning mechanism, and the labels are obtained from the determination of the residual performance after the material combustion experiment or expert manual scoring. The system uses the mean square error or Huber loss function as the optimization objective and jointly trains with other loss functions during the training process, so as to achieve end-to-end unified optimization between type classification and severity evaluation.

[0034] The above describes the method for identifying the surface defects of flame-retardant materials in the embodiments of the present invention. Next, the device for identifying the surface defects of flame-retardant materials in the embodiments of the present invention will be described. Please refer to Figure 2 , an embodiment of the device for identifying the surface defects of flame-retardant materials in the embodiments of the present invention includes: An acquisition module, configured to perform multi-angle image acquisition and preprocessing on the flame-retardant material combustion test sample to obtain a standardized multi-view image data set; A feature extraction module, which is used to perform multi-model feature extraction and feature matching processing on the standardized multi-view image dataset to obtain a set of material defect representation vectors and a set of defect semantic features; A calculation module, which is used to construct a spatial feature dual-constraint connection matrix based on the set of material defect representation vectors and the set of defect semantic features, and through material prior knowledge modulation and attention weight calculation, obtain an adaptive connection structure and a hierarchical defect map; A modeling module, which is used to input the adaptive connection structure and the hierarchical defect map into a deep graph neural network for multi-layer information transmission and explicit modeling of topological relationships, to obtain a set of deep representations of defect nodes and a material defect relationship matrix; An output module, which is used to perform domain-invariant feature extraction and structural consistency constraint on the set of deep representations of defect nodes and the material defect relationship matrix, to obtain a material-independent defect type distribution and a quantification result of defect severity.

[0035] Through the collaborative cooperation of the above-mentioned various components, the present invention can comprehensively capture the defect characteristics of each angle on the surface of the flame-retardant material, effectively solve the problem of uneven illumination on the surface of the flame-retardant material through the adaptive histogram equalization and illumination normalization algorithms, and ensure the accuracy and integrity of defect feature extraction. It effectively solves the problem of insufficient labeled data in the defect recognition of flame-retardant materials, and through the fusion of multiple visual model features and the professional text description of flame-retardant materials, realizes high-quality representation learning with a small amount of labeled data. By combining material prior knowledge and the attention mechanism, it adaptively adjusts the connection relationship between different defect nodes, effectively captures the topological dependence relationship between the defects of the flame-retardant material, and overcomes the limitation of traditional methods that only focus on the defect feature distribution. Through the residual jump connection and feature separation mechanism, it effectively alleviates the over-smoothing problem in the graph neural network, extracts the long-range interaction characteristics between defects, and enhances the discriminative ability of defect feature representation. Through domain adversarial training and structural consistency constraint, it solves the domain shift problem between different flame-retardant materials, enables the defect recognition method to have cross-material generalization ability, and realizes material-independent general defect recognition. It can adaptively process multi-modal defect data, comprehensively evaluate multi-dimensional indexes such as carbonization degree, structural integrity, and thermal stability, and provides a comprehensive and objective quantitative basis and material optimization direction for the performance evaluation of flame-retardant materials.

[0036] Refer to Figure 3 , in the embodiment of the present invention, a computer device is further provided. This computer device can be a server, and its internal structure can be as Figure 3As shown in the figure. The computer device includes a processor, a memory, a display screen, an input device, a network interface, and a database connected through a system bus. Among them, the processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0037] Those skilled in the art can understand that Figure 3 the structure shown in is only a block diagram of some structures related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0038] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0039] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium provided by the present invention and used in the embodiments can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or an external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0040] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the above-described systems, systems, and units can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0041] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0042] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for identifying surface defects of a flame retardant material, characterized in that, Including: Performing multi - angle image acquisition and pre - processing on the combustion test sample of the flame - retardant material to obtain a standardized multi - perspective image dataset; Performing multi - model feature extraction and feature matching processing on the standardized multi - perspective image dataset to obtain a set of material defect representation vectors and a set of defect semantic features; Constructing a spatial feature dual - constraint connection matrix based on the set of material defect representation vectors and the set of defect semantic features, and through material prior knowledge modulation and attention weight calculation, obtaining an adaptive connection structure and a hierarchical defect map; Inputting the adaptive connection structure and the hierarchical defect map into a deep graph neural network for multi - layer information transmission and explicit modeling of topological relationships, obtaining a set of deep representations of defect nodes and a material defect relationship matrix; Performing domain - invariant feature extraction and structural consistency constraint on the set of deep representations of defect nodes and the material defect relationship matrix, obtaining a material - independent defect type distribution and a quantification result of defect severity; 2. The method for identifying surface defects of a flame retardant material according to claim 1, characterized in that The performing multi - angle image acquisition and pre - processing on the combustion test sample of the flame - retardant material to obtain a standardized multi - perspective image dataset includes: Placing the combustion test sample of the flame - retardant material at the center of an anti - reflection turntable, and performing multi - angle image acquisition on the combustion test sample of the flame - retardant material through five high - resolution industrial cameras distributed at fixed angles to obtain original multi - angle image data; Performing adaptive histogram equalization processing on the original multi - angle image data to obtain a contrast - enhanced image, and performing bilateral filtering processing on the contrast - enhanced image to obtain a noise - reduced image; Performing image registration and geometric correction processing on the noise - reduced image to obtain a corrected image, and performing multi - scale segmentation processing on the corrected image to obtain a preliminary feature map; Performing illumination normalization processing on the preliminary feature map to obtain a standardized multi - perspective image dataset.

3. The method for identifying surface defects of a flame retardant material according to claim 1, wherein The performing multi - model feature extraction and feature matching processing on the standardized multi - perspective image dataset to obtain a set of material defect representation vectors and a set of defect semantic features includes: Inputting the standardized multi - perspective image dataset into multiple pre - trained contrastive learning vision - language models for image block analysis to obtain preliminary image feature data; Constructing a set of semantic descriptions of flame - retardant material defects according to the professional knowledge of flame - retardant material defects, and performing text feature encoding on the set of semantic descriptions of flame - retardant material defects to obtain a set of defect semantic features; Calculating the similarity between the preliminary image feature data and the set of defect semantic features to obtain a feature similarity distribution matrix; Constructing a contrastive learning objective function based on the feature similarity distribution matrix, and by minimizing the intra - class sample distance and maintaining the inter - class boundary interval, obtaining an optimized feature space; Inputting the image features in the optimized feature space into a feature transformation network to perform non - linear transformation and block - dropping regularization processing to obtain a set of material defect representation vectors.

4. The method for identifying surface defects of a flame retardant material according to claim 1, wherein The constructing a spatial feature dual - constraint connection matrix based on the set of material defect representation vectors and the set of defect semantic features, and through material prior knowledge modulation and attention weight calculation, obtaining an adaptive connection structure and a hierarchical defect map includes: Construct a spatial feature double-constraint connection matrix based on the spatial positions and characteristic distances of the vectors in the material defect representation vector set; Input the material defect representation vector set into a linear transformation module to calculate a query matrix and a key matrix, and obtain a defect self-attention matrix; Construct a material prior knowledge matrix based on the defect semantic feature set; Perform an element-wise product operation on the defect self-attention matrix and the material prior knowledge matrix and perform a non-linear activation process to obtain a connection modulation weight matrix; Perform an element-wise product operation on the spatial feature double-constraint connection matrix and the connection modulation weight matrix to obtain an adaptive connection structure; Cluster the material defect representation vector set hierarchically according to feature similarity and calculate the inter-layer connection strength to obtain a hierarchical defect map; 5. The method for identifying surface defects of a flame retardant material according to claim 4, wherein The step of clustering the material defect representation vector set hierarchically according to feature similarity and calculating the inter-layer connection strength to obtain a hierarchical defect map includes: Calculate the Euclidean distance of each vector in the material defect representation vector set to obtain a feature similarity matrix; Perform spectral clustering analysis on the material defect representation vector set based on the feature similarity matrix to obtain four layers of defect feature clusters, and the feature dimensions of the four layers of defect feature clusters are 64, 128, 256, and 512 respectively; Construct an adjacency relationship for the nodes within each layer of defect feature clusters to obtain four intra-layer connection structure diagrams; Calculate the connection weights between the four layers of defect feature clusters according to the feature correlation between cross-layer nodes to obtain an inter-layer connection matrix; Integrate the intra-layer connection structure diagrams and the inter-layer connection matrix to obtain an initial hierarchical graph structure; Adopt a self-attention mechanism to optimize the inter-layer information flow of the initial hierarchical graph structure to obtain a hierarchical defect map; 6. The method for identifying surface defects of a flame retardant material according to claim 1, characterized in that, The step of inputting the adaptive connection structure and the hierarchical defect map into a deep graph neural network for multi-layer information transmission and explicit modeling of topological relationships to obtain a set of deep representations of defect nodes and a material defect relationship matrix includes: Input the adaptive connection structure and the hierarchical defect map into the deep node graph representation learning module of the deep graph neural network to extract node features and obtain initial multi-layer node features; Calculate the attention scores for the neighborhood of each node based on the adaptive connection structure, and multiply by the connection weights through a bilinear attention mechanism activated by LeakyReLU to obtain neighborhood information transmission weights; Input the initial multi-layer node features into the feature transfer network of the deep graph neural network for feature fusion to obtain multi-layer fusion features; Input the multi-layer fusion features into a feature decoupling network for gradient projection to obtain a set of deep representations of defect nodes; Input the set of deep representations of defect nodes into a graph projection network for relationship analysis to obtain an initial matrix of defect type relationship strengths; Calculate the Frobenius norm difference between the initial matrix of defect type relationship strengths and an ideal relationship matrix constructed based on flame retardant material science knowledge, and optimize it through gradient descent iteration to obtain a material defect relationship matrix.

7. The method for identifying surface defects of a flame retardant material according to claim 1, wherein Performing domain-invariant feature extraction and structural consistency constraint on the defect node depth representation set and the material defect relationship matrix to obtain a material-independent defect type distribution and a defect severity quantification result, including: Inputting the defect node depth representation set into a gradient reversal layer to construct a domain classifier and performing minimization of domain classification cross-entropy loss calculation to obtain a domain-invariant feature representation; Calculating the Laplacian matrix and the connection matrix of the source domain and the target domain according to the defect node depth representation set and the material defect relationship matrix, and obtaining a structural consistency constraint representation by minimizing the matrix norm difference; Calculating the inter-domain similarity weight based on the attribute similarity of different flame retardant material types, and performing weighted norm calculation on the material defect relationship matrix according to the inter-domain similarity weight to obtain a multi-domain adversarial loss value; Inputting the domain-invariant feature representation into a graph convolutional classifier for non-linear mapping to obtain a material-independent defect type distribution; Inputting the domain-invariant feature representation and the structural consistency constraint representation into a regression network for severity score calculation to obtain a defect severity quantification result.

8. An image recognition device for surface defects of a flame retardant material, characterized in that, For performing the flame retardant material surface defect image recognition method according to any one of claims 1-7, the flame retardant material surface defect image recognition device includes: An acquisition module, configured to perform multi-angle image acquisition and preprocessing on a flame retardant material combustion test sample to obtain a standardized multi-view image data set; A feature extraction module, configured to perform multi-model feature extraction and feature matching processing on the standardized multi-view image data set to obtain a material defect representation vector set and a defect semantic feature set; A calculation module, configured to construct a spatial feature double-constraint connection matrix based on the material defect representation vector set and the defect semantic feature set, and obtain an adaptive connection structure and a hierarchical defect map through material prior knowledge modulation and attention weight calculation; A modeling module, configured to input the adaptive connection structure and the hierarchical defect map into a deep graph neural network for multi-layer information transmission and explicit modeling of topological relationships to obtain a defect node depth representation set and a material defect relationship matrix; An output module, configured to perform domain-invariant feature extraction and structural consistency constraint on the defect node depth representation set and the material defect relationship matrix to obtain a material-independent defect type distribution and a defect severity quantification result.

9. A computer device, characterized in that, Including a memory and a processor, the memory stores a computer program that can run on the processor, and when the processor executes the computer program, it implements the flame retardant material surface defect image recognition method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, On which a computer program is stored, and when the computer program is run by the processor, the processor is caused to execute the flame retardant material surface defect image recognition method according to any one of claims 1 to 7.

Images