A deep learning-based injection molded product surface image defect recognition method

Abstract

The application discloses a kind of based on deep learning's injection molding product surface image defect identification method, including the following steps: injection molding product surface image data is collected, generates standardization surface image dataset;Defect candidate region extraction processing is executed, and defect candidate region set is generated;Defect structure chart is constructed;Action feature vector is constructed, and action behavior representation processing is executed, and defect action mode set is generated;Improved VGAE is constructed, and action mode collaborative hidden structure is generated by action mode hidden structure generation layer;Collaborative reconstruction processing is executed, and action mode collaborative reconstruction diagram is generated;Collaborative difference analysis is executed, and abnormal action mode collaborative structure is generated;Injection molding product surface image defect identification result is generated.The defect action mode collaborative analysis method of the present application using improved VGAE realizes injection molding product surface defect intelligent identification, with the advantages of high recognition accuracy, strong collaborative representation ability and good stability.

Description

Technical Field

[0001] This invention relates to the field of surface image defect recognition of injection molded products, and more particularly to a method for surface image defect recognition of injection molded products based on deep learning. Background Technology

[0002] Injection molded products are widely used in automotive parts, electronic product housings, home appliance components, and industrial products. Their surface quality directly affects product performance, assembly accuracy, and market value. In existing technologies, manual visual inspection or deep learning-based image recognition methods are commonly used to detect surface defects in injection molded products. By extracting texture features, edge features, or region features, defects such as shrinkage marks, cracks, bubbles, and deformation can be identified, thereby achieving product quality inspection and screening.

[0003] However, most existing defect identification methods focus on the local feature analysis of a single defect region or the direct correlation modeling between defect regions. They are difficult to effectively characterize the common change behavior patterns and propagation correlation patterns among multiple defect regions. They lack the ability to explore the potential collaborative relationships between different defect regions. When the defect formation process has complex collaborative evolution characteristics, problems such as incomplete defect structure expression, insufficient ability to recover correlation relationships, and decreased accuracy of abnormal pattern recognition may occur, thus affecting the accuracy and stability of the defect identification results of injection molded product surface images. Summary of the Invention

[0004] One objective of this invention is to propose a method for identifying surface defects in injection molded products based on deep learning. This invention employs an improved VGAE defect action mode collaborative analysis method to achieve intelligent identification of surface defects in injection molded products, which has the advantages of high identification accuracy, strong collaborative characterization ability, and good stability.

[0005] A method for identifying surface image defects of injection-molded products based on deep learning according to an embodiment of the present invention includes the following steps:

[0006] Collect surface image data of injection molded products and preprocess them to generate a standardized surface image dataset;

[0007] Perform defect candidate region extraction processing on a standardized surface image dataset to generate a set of defect candidate regions;

[0008] A defect structure graph is constructed based on the relationship between multiple defect candidate regions. The defect candidate regions are mapped as graph nodes, and the relationship between the candidate regions is mapped as graph edges to generate the defect structure graph.

[0009] Based on the graph node features and graph edge association features in the defect structure graph, an action feature vector is constructed. Action behavior representation processing is performed on multiple action feature vectors to generate a set of defect action patterns.

[0010] An improved VGAE is constructed, which includes an action mode implicit structure generation layer and an action mode cooperative reconstruction layer. The defect structure diagram and the defect action mode set are input into the improved VGAE, and the action mode cooperative implicit structure is generated through the action mode implicit structure generation layer.

[0011] The collaborative reconstruction layer of action mode performs collaborative reconstruction processing based on the implicit structure of action mode collaboration, and generates an action mode collaborative reconstruction graph.

[0012] Perform a collaborative difference analysis on the collaborative reconstruction diagram of action modes and the defect structure diagram to generate a collaborative structure of abnormal action modes;

[0013] Based on the abnormal action mode, the surface image defect recognition results of injection molded products are generated.

[0014] Optionally, the surface image data specifically includes surface texture images, surface edge images, and surface grayscale images; the preprocessing specifically includes image denoising, image size unification, image brightness correction, and image enhancement.

[0015] Optionally, the generation of the defect candidate region set specifically includes:

[0016] Texture anomaly region extraction is performed on surface texture images in the standardized surface image dataset to generate texture anomaly regions; edge anomaly region extraction is performed on surface edge images to generate edge anomaly regions; structural anomaly region extraction is performed on surface grayscale images to generate structural anomaly regions; region correspondence analysis is performed on texture anomaly regions, edge anomaly regions, and structural anomaly regions to generate candidate defect regions; and region organization processing is performed on multiple candidate defect regions to generate a set of candidate defect regions.

[0017] Optionally, the generation of the defect structure diagram specifically includes:

[0018] Extract multiple candidate defect regions from the defect candidate region set, and determine the spatial location information, texture feature information, edge feature information, and structural feature information corresponding to each candidate defect region; generate candidate association relationships based on the spatial adjacency relationship, texture similarity relationship, edge continuity relationship, and structural association relationship between multiple candidate defect regions; map multiple candidate defect regions to corresponding graph nodes, and map multiple candidate association relationships to corresponding graph edges to generate a defect structure graph.

[0019] Optionally, the generation of the defect action mode set specifically includes:

[0020] Read multiple graph nodes and multiple graph edges in the defect structure graph, extract the graph node features corresponding to each graph node and the graph edge association features corresponding to each graph edge, and generate a feature data set.

[0021] Based on the graph node features and graph edge association features in the feature data set, construct corresponding action feature vectors and generate an action feature vector set;

[0022] Perform change consistency analysis and propagation synergy analysis on multiple action feature vectors in the action feature vector set to generate action behavior representation results;

[0023] Based on the behavioral representation results, behavioral aggregation is performed on the behavioral feature vectors with the same behavioral characteristics to generate multiple defect behavioral patterns;

[0024] Perform pattern organization processing on multiple defect action patterns to generate a defect action pattern set.

[0025] Optionally, the construction of the improved VGAE specifically includes:

[0026] Extract multiple defect action patterns from the defect action pattern set, establish action pattern nodes based on the correspondence between the action behaviors of multiple defect action patterns, and generate an action pattern node set.

[0027] Around multiple action pattern nodes in the action pattern node set, extract the collaborative change state and propagation association state between different action pattern nodes, establish the pattern connection relationship between action pattern nodes, and generate the action pattern connection structure.

[0028] Based on the set of action pattern nodes and the action pattern connection structure, an action pattern hidden structure generation layer is constructed. The action pattern nodes are used as the hidden structure learning objects, and the pattern connection relationship is used as the hidden structure propagation path to generate the action pattern hidden structure generation layer.

[0029] Read the set of action mode nodes and the action mode connection structure, extract the collaborative association state and collaborative propagation state between each action mode node, construct action mode collaborative reconstruction rules, and generate a set of collaborative reconstruction rules.

[0030] Based on the set of action mode nodes, the action mode connection structure, and the set of collaborative reconstruction rules, an action mode collaborative reconstruction layer is constructed, and the collaborative relationship between action mode nodes is used as the reconstruction object to generate the action mode collaborative reconstruction layer.

[0031] Establish the data transfer relationship between the implicit structure generation layer of the action mode and the collaborative reconstruction layer of the action mode, integrate the implicit structure generation layer of the action mode and the collaborative reconstruction layer of the action mode, and generate an improved VGAE.

[0032] Optionally, the generation of the cooperative implicit structure of the action mode specifically includes:

[0033] Map multiple graph nodes in the defect structure graph to multiple defect action modes in the defect action mode set to generate action mode mapping results;

[0034] Based on the effect mode mapping results, the graph nodes corresponding to the same defect effect mode are merged to form multiple effect mode nodes, and the mode connection relationship between the multiple effect mode nodes is established.

[0035] Input multiple action mode nodes and mode connection relationships into the action mode implicit structure generation layer in the improved VGAE to generate the action mode propagation structure.

[0036] The implicit structure generation layer of the action pattern is used to perform implicit space propagation processing on multiple action pattern nodes in the action pattern propagation structure to generate action pattern collaborative representation results.

[0037] Based on the collaborative characterization results of action modes, the latent space relationships between nodes of different action modes are extracted to generate the action mode association structure.

[0038] Perform structural organization processing on multiple implicit space relationships in the action mode association structure to generate an action mode cooperative implicit structure.

[0039] Optionally, the generation of the collaborative reconstruction graph of the action mode specifically includes:

[0040] Read multiple action pattern nodes and multiple latent space relationships in the collaborative implicit structure of action patterns, and generate collaborative reconstruction input results;

[0041] The collaborative reconstruction input results are input into the action mode collaborative reconstruction layer, and collaborative reconstruction processing is performed on the latent space association relationships between multiple action mode nodes to generate action mode collaborative reconstruction results.

[0042] Based on the collaborative reconstruction results of the action mode, the collaborative association state between multiple action mode nodes is restored, and the collaborative association structure of the action mode is generated.

[0043] Perform graph structure mapping processing on the action mode collaborative association structure to generate an action mode collaborative reconstruction graph.

[0044] Optionally, the generation of the abnormal action mode cooperative structure specifically includes:

[0045] Read multiple reconstructed graph nodes and edges from the collaborative reconstructed graph of the action mode, and read multiple graph nodes and edges from the defect structure graph to establish the correspondence between the collaborative reconstructed graph of the action mode and the defect structure graph, and generate the graph structure correspondence result.

[0046] Based on the graph structure correspondence results, node difference analysis is performed on multiple reconstructed graph nodes and their corresponding graph nodes to extract the feature deviation states between the corresponding nodes and generate node difference results;

[0047] Perform edge association difference analysis on multiple reconstructed graph edges and their corresponding graph edges, extract the association deviation state between corresponding graph edges, and generate edge difference results;

[0048] The node difference results and edge difference results are fused together to generate a collaborative difference result;

[0049] Based on the results of collaborative differences, the reconstructed graph nodes with collaborative deviation states and their corresponding reconstructed graph edges are located, and a set of nodes with abnormal action patterns and a set of edges associated with abnormal action patterns are generated.

[0050] Based on the set of nodes and edges associated with the abnormal action patterns, structural organization processing is performed to generate a collaborative structure for the abnormal action patterns.

[0051] Optionally, the generation of the surface image defect recognition result of the injection molded product specifically includes:

[0052] Read multiple abnormal action mode nodes and multiple abnormal action mode associated edges in the abnormal action mode collaborative structure, establish the mapping relationship between abnormal action mode nodes and corresponding defect candidate regions, and generate abnormal region results.

[0053] Based on the abnormal region results, extract the texture change features, edge change features, and structural change features from the corresponding defect candidate regions to generate abnormal feature results;

[0054] Based on the abnormal feature results and the collaborative deviation state corresponding to the associated edge of the abnormal action mode, defect category determination processing is performed to generate defect category results;

[0055] The defect location result is generated by combining the defect category result with the location distribution status of the corresponding defect candidate region.

[0056] Based on the defect category and defect location results, generate defect recognition results for the surface image of the injection molded product.

[0057] The beneficial effects of this invention are:

[0058] This invention proposes a deep learning-based method for identifying surface defects in injection-molded products. By constructing a defect structure map, a set of defect action patterns, and an improved VGAE network, it achieves in-depth mining of the formation and co-evolution patterns of surface defects in injection-molded products. Unlike existing technologies that mainly rely on local features of defect regions or simple regional relationships for identification, this invention first constructs a defect structure map based on the spatial adjacency, texture similarity, edge continuity, and structural relationships among multiple defect candidate regions. Furthermore, it extracts the common change and propagation behavior patterns of multiple defect candidate regions to generate a set of defect action patterns. This enables the defect identification process to not only focus on the abnormal features of a single defect region itself, but also on the co-change and propagation relationships among multiple defect regions, thereby improving the feature representation capability in complex defect scenarios.

[0059] This invention constructs an improved VGAE comprising an action pattern implicit structure generation layer and an action pattern collaborative reconstruction layer. It elevates the network learning object from traditional graph nodes to the defect action pattern level. Through action pattern connection structures, it establishes an information propagation mechanism between nodes of different action patterns, learns explicit and implicit collaborative relationships in the latent space, generates action pattern collaborative implicit structures, and further restores the collaborative association states between nodes of different action patterns, forming an action pattern collaborative reconstruction graph. By performing collaborative difference analysis on the action pattern collaborative reconstruction graph and the defect structure graph, it can effectively discover potential abnormal action patterns and abnormal propagation relationships that are difficult to identify using traditional structural association analysis, thereby improving the ability to locate abnormal defects and the accuracy of defect identification. Because this invention can simultaneously utilize local feature information, structural association information, and action pattern collaborative information of the defect region, it can more completely describe the defect formation and propagation process, achieving high defect identification accuracy, strong abnormal pattern mining ability, good stability of identification results, and strong robustness. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of a deep learning-based method for identifying surface image defects in injection molded products, as proposed in this invention.

[0062] Figure 2 This is a schematic diagram of the overall structure of the improved VGAE for a deep learning-based surface image defect recognition method for injection molded products proposed in this invention.

[0063] Figure 3This is a schematic diagram of the hidden structure generation layer structure of the deep learning-based surface image defect recognition method for injection molded products proposed in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0065] refer to Figures 1-3 A method for identifying surface image defects in injection molded products based on deep learning, comprising the following steps:

[0066] Collect surface image data of injection molded products and preprocess them to generate a standardized surface image dataset;

[0067] Perform defect candidate region extraction processing on a standardized surface image dataset to generate a set of defect candidate regions;

[0068] A defect structure graph is constructed based on the relationship between multiple defect candidate regions. The defect candidate regions are mapped as graph nodes, and the relationship between the candidate regions is mapped as graph edges to generate the defect structure graph.

[0069] Based on the graph node features and graph edge association features in the defect structure graph, an action feature vector is constructed. Action behavior representation processing is performed on multiple action feature vectors to generate a set of defect action patterns.

[0070] An improved VGAE is constructed, which includes an action mode implicit structure generation layer and an action mode cooperative reconstruction layer. The defect structure diagram and the defect action mode set are input into the improved VGAE, and the action mode cooperative implicit structure is generated through the action mode implicit structure generation layer.

[0071] The collaborative reconstruction layer of action mode performs collaborative reconstruction processing based on the implicit structure of action mode collaboration, and generates an action mode collaborative reconstruction graph.

[0072] Perform a collaborative difference analysis on the collaborative reconstruction diagram of action modes and the defect structure diagram to generate a collaborative structure of abnormal action modes;

[0073] Based on the abnormal action mode, the surface image defect recognition results of injection molded products are generated.

[0074] In this embodiment, the surface image data specifically includes surface texture image, surface edge image, and surface grayscale image; the preprocessing specifically includes image denoising processing, image size unification processing, image brightness correction processing, and image enhancement processing.

[0075] In this embodiment, the generation of the defect candidate region set specifically includes:

[0076] Texture anomaly region extraction is performed on surface texture images in the standardized surface image dataset to generate texture anomaly regions; edge anomaly region extraction is performed on surface edge images to generate edge anomaly regions; structural anomaly region extraction is performed on surface grayscale images to generate structural anomaly regions; region correspondence analysis is performed on texture anomaly regions, edge anomaly regions, and structural anomaly regions to generate candidate defect regions; region organization processing is performed on multiple candidate defect regions to generate a set of candidate defect regions.

[0077] When performing structural anomaly region extraction processing on a surface grayscale image, the grayscale values ​​and grayscale distribution of each pixel in the surface grayscale image are first read, and the grayscale changes and grayscale gradient changes between different regions are extracted. Since structural defects such as shrinkage marks, depressions, cracks, bubbles, and deformations on the surface of injection molded products will change the surface morphology of the corresponding areas, thereby causing changes in the grayscale distribution and grayscale change states of local areas, the corresponding structural anomaly regions are identified based on the grayscale distribution anomaly regions, grayscale change anomaly regions, and grayscale gradient anomaly regions in the surface grayscale image; structural anomaly regions are used to characterize areas where abnormal changes in surface morphology occur.

[0078] In this embodiment, the generation of the defect structure diagram specifically includes:

[0079] Extract multiple defect candidate regions from the defect candidate region set, and determine the spatial location information, texture feature information, edge feature information, and structural feature information corresponding to each defect candidate region; generate candidate association relationships based on the spatial adjacency relationship, texture similarity relationship, edge continuity relationship, and structural association relationship between multiple defect candidate regions; map multiple defect candidate regions to corresponding graph nodes, and map multiple candidate association relationships to corresponding graph edges to generate a defect structure graph;

[0080] Spatial location information is used to characterize the positional distribution of the defect candidate region in the surface image of the injection molded product; texture feature information is used to characterize the texture change state inside the defect candidate region; edge feature information is used to characterize the edge change state of the boundary of the defect candidate region; structural feature information is used to characterize the gray-scale distribution state and structural change state inside the defect candidate region.

[0081] Based on the spatial location information corresponding to different defect candidate regions, spatial adjacency relationships are established. These relationships characterize the proximity of two defect candidate regions and their spatial distribution association. Based on the texture feature information corresponding to different defect candidate regions, texture similarity relationships are established. These relationships characterize the consistency of texture changes and the proximity of texture distribution between two defect candidate regions. Based on the edge feature information corresponding to different defect candidate regions, edge continuity relationships are established. These relationships characterize the continuity of edge changes and the extension of boundaries between two defect candidate regions. Based on the structural feature information corresponding to different defect candidate regions, structural association relationships are established. These relationships characterize the association of structural changes and the proximity of structural evolution between two defect candidate regions.

[0082] After obtaining spatial adjacency, texture similarity, edge continuity, and structural association, the region combinations between different defect candidate regions are traversed sequentially. The spatial adjacency, texture similarity, edge continuity, and structural association of each defect candidate region pair are statistically analyzed. Multiple relationships belonging to the same defect candidate region pair are recorded in a unified manner to form the candidate association relationship of the corresponding defect candidate region pair. Each candidate association relationship is used to record the positional association state, texture association state, edge association state, and structural association state between the corresponding defect candidate regions.

[0083] In this embodiment, the generation of the defect action mode set specifically includes:

[0084] Read multiple graph nodes and multiple graph edges in the defect structure graph, extract the graph node features corresponding to each graph node and the graph edge association features corresponding to each graph edge, and generate a feature data set.

[0085] Based on the graph node features and graph edge association features in the feature data set, construct corresponding action feature vectors and generate an action feature vector set;

[0086] The generation of the feature vector set specifically includes:

[0087] This process reads multiple graph node features and multiple graph edge association features from a feature dataset. It then performs feature processing on the texture, edge, and structural features corresponding to each graph node feature to generate node feature results. Based on the candidate associations corresponding to each graph node, it extracts spatial adjacency, texture similarity, edge continuity, and structural associations from the corresponding graph edge association features to generate edge association feature results. Next, it performs feature correspondence processing on the node feature results and the edge association feature results connected to each graph node to establish a feature correspondence between the node feature results and the edge association feature results, generating feature correspondence results. Based on the feature correspondence results, it performs vector organization processing on the node feature results and the edge association feature results to generate the corresponding graph node's impact feature vector. Finally, it performs vector aggregation processing on the impact feature vectors corresponding to multiple graph nodes to generate an impact feature vector set.

[0088] The action feature vector is used to characterize the self-change features and association propagation features of the corresponding graph node in the defect structure graph. The self-change features include the texture feature information, edge feature information and structural feature information of the corresponding defect candidate region. The association propagation features include the spatial adjacency relationship, texture similarity relationship, edge continuity relationship and structural association relationship between the corresponding graph node and its neighboring graph nodes. Each graph node generates an action feature vector, and the action feature vectors corresponding to multiple graph nodes together constitute the action feature vector set.

[0089] Perform change consistency analysis and propagation synergy analysis on multiple action feature vectors in the action feature vector set to generate action behavior representation results;

[0090] Change consistency analysis is used to analyze whether the self-change features corresponding to different action feature vectors have a consistent change trend. The self-change features include the texture feature information, edge feature information, and structural feature information of the corresponding graph node. When the texture change state, edge change state, and structural change state corresponding to different action feature vectors show the same change trend, it is determined that there is a change consistency state between the corresponding action feature vectors.

[0091] Propagation synergy analysis is used to analyze whether the associated propagation features corresponding to different action feature vectors have a synergistic change state. The associated propagation features are derived from the positional association state, texture association state, edge association state, and structural association state recorded by the graph edge association features. When the associated propagation features corresponding to different action feature vectors show the same propagation trend, it is determined that there is a propagation synergy state between the corresponding action feature vectors.

[0092] After completing the change consistency analysis and propagation synergy analysis, the change consistency state and propagation synergy state of the corresponding action feature vectors are uniformly organized to form the action behavior representation result of the corresponding action feature vectors; the action behavior representation result is used to characterize the change behavior state and propagation behavior state of the corresponding action feature vectors in the defect structure diagram;

[0093] Change consistency analysis specifically includes:

[0094] Based on the texture, edge, and structural features in multiple action feature vectors, the feature change state corresponding to each action feature vector is determined. Based on the feature change states corresponding to different action feature vectors, action feature vectors with the same texture change direction, edge change direction, and structural change direction are identified. Action feature vectors with the same feature change direction are grouped into the same change set, forming a consistent change state for the corresponding action feature vectors. Based on the consistent change state, a consistency analysis result for the corresponding action feature vectors is generated.

[0095] The analysis of propagation synergy specifically includes:

[0096] The process involves reading the associated propagation features corresponding to multiple action feature vectors, including location association state, texture association state, edge association state, and structural association state. It then performs association state comparison processing on the association propagation features corresponding to different action feature vectors to determine the changes in location association state, texture association state, edge association state, and structural association state for each action feature vector. Based on the changes in association state corresponding to different action feature vectors, it performs propagation trend analysis to identify action feature vectors with the same trend in association state change. For multiple action feature vectors with the same trend in association state change, it performs collaborative classification processing to generate propagation collaborative states. Finally, it generates the propagation collaborative analysis results for the corresponding action feature vectors based on the propagation collaborative states.

[0097] Based on the behavioral representation results, behavioral aggregation is performed on the behavioral feature vectors with the same behavioral characteristics to generate multiple defect behavioral patterns;

[0098] The generation of defect action modes specifically includes:

[0099] The system iterates through multiple behavioral representations and extracts the change behavior state and propagation behavior state corresponding to each feature vector. Feature vectors with the same change behavior state and propagation behavior state are grouped into the same behavior category to form a behavior aggregation set corresponding to the feature vectors. For each behavior aggregation set, the change behavior state and propagation behavior state jointly expressed by each feature vector are uniformly organized to generate a corresponding defect action pattern. The defect action pattern is used to characterize the change behavior pattern and propagation behavior pattern jointly expressed by multiple defect candidate regions.

[0100] Perform pattern organization processing on multiple defect action patterns to generate a defect action pattern set.

[0101] In this embodiment, constructing the improved VGAE specifically includes:

[0102] Extract multiple defect action patterns from the defect action pattern set, establish action pattern nodes based on the correspondence between the action behaviors of multiple defect action patterns, and generate an action pattern node set.

[0103] Around multiple action pattern nodes in the action pattern node set, extract the collaborative change state and propagation association state between different action pattern nodes, establish the pattern connection relationship between action pattern nodes, and generate the action pattern connection structure.

[0104] The generation of the action mode connection structure specifically includes:

[0105] Read multiple action pattern nodes from the action pattern node set; extract the defect action pattern corresponding to each action pattern node; compare the changes and propagation of the defect action patterns corresponding to different action pattern nodes; identify action pattern nodes with the same changes or propagation as associated nodes; establish pattern connection relationships between associated nodes; generate an action pattern connection structure based on multiple pattern connection relationships; the action pattern connection structure is used to record the connection status between multiple action pattern nodes; the action pattern connection structure includes multiple action pattern nodes and their corresponding pattern connection relationships.

[0106] Based on the set of action pattern nodes and the action pattern connection structure, an action pattern hidden structure generation layer is constructed. The action pattern nodes are used as the hidden structure learning objects, and the pattern connection relationship is used as the hidden structure propagation path to generate the action pattern hidden structure generation layer.

[0107] The construction of the implicit structure generation layer of the action mode specifically includes:

[0108] The process involves reading multiple action pattern nodes from the action pattern node set and the pattern connection relationships in the action pattern connection structure; using multiple action pattern nodes as latent structure learning inputs and pattern connection relationships as information transmission paths between different action pattern nodes; determining other action pattern nodes connected to each action pattern node based on the pattern connection relationships and transmitting the information of the corresponding connected action pattern nodes to the current action pattern node; summarizing and processing the information of the current action pattern node and the information transmitted by the corresponding connected action pattern nodes to generate the node representation result corresponding to the current action pattern node; processing multiple action pattern nodes separately to generate the node representation result corresponding to each action pattern node; and forming the action pattern latent structure generation layer based on the node representation results corresponding to multiple action pattern nodes.

[0109] The action mode node information includes the defect action mode associated with the corresponding action mode node and the graph node information that forms the action mode node; the node representation result is used to record the information of the current action mode node and the information of other action mode nodes connected to the current action mode node; the node representation result includes the defect action mode corresponding to the current action mode node and the defect action modes corresponding to other action mode nodes obtained through the mode connection relationship.

[0110] Read the set of action mode nodes and the action mode connection structure, extract the collaborative association state and collaborative propagation state between each action mode node, construct action mode collaborative reconstruction rules, and generate a set of collaborative reconstruction rules.

[0111] The generation of the collaborative reconstruction rule set specifically includes:

[0112] Read multiple action pattern nodes from the action pattern node set and the pattern connection relationships in the action pattern connection structure; determine the other action pattern nodes connected to each action pattern node based on the pattern connection relationships; record the other action pattern nodes connected to each action pattern node; use the correspondence between each action pattern node and its corresponding connected action pattern nodes as action pattern collaborative reconstruction rules; perform unified organization processing on multiple action pattern collaborative reconstruction rules to generate a collaborative reconstruction rule set; action pattern collaborative reconstruction rules are used to record other action pattern nodes that each action pattern node needs to refer to during the collaborative reconstruction process; each rule corresponds to one action pattern node and its associated other action pattern nodes, forming a reference correspondence between nodes;

[0113] Based on the set of action mode nodes, the action mode connection structure, and the set of collaborative reconstruction rules, an action mode collaborative reconstruction layer is constructed, and the collaborative relationship between action mode nodes is used as the reconstruction object to generate the action mode collaborative reconstruction layer.

[0114] Read the set of action mode nodes and the collaborative relationships in the action mode connection structure; based on the set of collaborative reconstruction rules, perform collaborative recovery processing on the explicit and implicit collaborative relationships between different action mode nodes to reconstruct the collaborative relationship state between the corresponding action mode nodes; during the recovery process, continuously maintain the collaborative correspondence state between the explicit and implicit collaborative relationships to generate the reconstructed collaborative relationship; based on the reconstructed collaborative relationship, establish the collaborative reconstruction structure between different action mode nodes to form the action mode collaborative reconstruction layer;

[0115] The collaborative association relationship is composed of explicit collaborative relationships represented by pattern connection relationships and implicit collaborative relationships represented by potential collaborative association relationships; the collaborative reconstruction rule is used to constrain the recovery process of collaborative association relationships between nodes with different action patterns and to record the collaborative correspondence state between explicit and implicit collaborative relationships; the action pattern collaborative reconstruction layer is used to recover the collaborative association relationships between nodes with different action patterns.

[0116] Establish the data transfer relationship between the implicit structure generation layer of the action mode and the collaborative reconstruction layer of the action mode, integrate the implicit structure generation layer of the action mode and the collaborative reconstruction layer of the action mode, and generate an improved VGAE.

[0117] The improvements in the improved VGAE specifically include:

[0118] I. Regarding the learning object, traditional VGAE typically uses graph nodes and their connections as the learning object. It learns graph structure representation by learning the features of the graph nodes and their connections. In the scenario of surface defect recognition for injection molded products, it mainly focuses on the local relationships between candidate defect regions, making it difficult to characterize the common change and propagation behavior patterns exhibited by different candidate defect regions. Improved VGAE transforms the learning object from graph nodes in the defect structure graph to action pattern nodes corresponding to defect action patterns. These action pattern nodes characterize the common change and propagation behavior patterns exhibited by multiple candidate defect regions, shifting the network's learning focus from the candidate defect region level to the defect action pattern level. This allows for the description of the intrinsic relationships between different candidate defect regions at the action pattern level, improving the expressive power of change and propagation behavior patterns. II. Regarding the information propagation mechanism, traditional VGAE mainly relies on the node connections in the graph structure for information propagation. The propagation process depends heavily on the direct connection states between graph nodes, and the propagated content primarily originates from the graph node's own features and its adjacency graph. Node characteristics; Improved VGAE establishes an information propagation mechanism between action mode nodes through an action mode connection structure, forming mode connection relationships based on collaborative change states and propagation association states, and using these connection relationships as information propagation paths. This allows change behavior information and propagation behavior information to be propagated and updated between different action mode nodes, enabling the network to learn the collaborative change states and propagation association states between different defect action modes, thus improving the modeling ability of change behavior patterns and propagation behavior patterns; Third, in terms of latent structure learning, traditional VGAE mainly learns the potential node association relationships between graph nodes, and its latent space representation reflects more the structural association features between graph nodes, with limited ability to express the latent space collaborative states between different action modes; Improved VGAE constructs an action mode latent structure generation layer, integrating collaborative change states and propagation association states during the propagation and update process of action mode nodes, generating latent space representation results for corresponding action mode nodes, and establishing potential collaborative association relationships between different action mode nodes based on multiple latent space representation results, forming action mode collaborative latent structure representations.By learning explicit cooperative relationships represented by pattern connections and implicit cooperative relationships represented by potential cooperative associations, the network can discover latent cooperative states between nodes of different action patterns that are not directly represented by pattern connections, thereby improving the ability to mine cooperative relationships of action patterns. Fourth, in terms of reconstruction objectives, the reconstruction process of traditional VGAE mainly focuses on restoring the connection relationships between graph nodes, and its reconstruction results reflect more the connection state of the graph structure itself. The improved VGAE constructs an action pattern cooperative reconstruction layer, taking the cooperative association relationships between action pattern nodes as the reconstruction object, and restoring the explicit and implicit cooperative relationships in the cooperative association relationships according to the cooperative reconstruction rule set, realizing the reconstruction of the cooperative association states between nodes of different action patterns. Compared with the traditional graph node connection relationship reconstruction method, this invention can further restore the cooperative association relationships between nodes of different action patterns, improving the integrity and stability of defective action pattern expression.

[0119] The improved VGAE includes an action pattern latent structure generation layer and an action pattern cooperative reconstruction layer. The action pattern latent structure generation layer, located at the network front end, receives the set of action pattern nodes and the action pattern connection structure. It performs latent space learning processing on the change behavior patterns and propagation behavior patterns corresponding to different action pattern nodes, generating latent space representation results for the corresponding action pattern nodes. Based on the latent space representation results, it establishes potential cooperative associations between different action pattern nodes, forming action pattern cooperative latent structure representations. The action pattern cooperative reconstruction layer, located after the action pattern latent structure generation layer, receives the action pattern cooperative latent structure representations output by the generation layer and, combined with the cooperative reconstruction rule set, performs latent space learning processing on the latent space representations of different action pattern nodes. The collaborative relationships between nodes of different action modes are restored, reconstructing both explicit and implicit collaborative relationships. During the inter-layer connection process, the latent space representation results, potential collaborative relationships, and action mode collaborative latent structure representations output by the action mode latent structure generation layer serve as input data for the action mode collaborative reconstruction layer. The action mode collaborative reconstruction layer performs collaborative reconstruction processing on the input data according to the collaborative reconstruction rule set, thereby forming a complete action mode collaborative reconstruction structure. By establishing the data transfer relationship between the action mode latent structure generation layer and the action mode collaborative reconstruction layer, the improved VGAE can complete the action mode collaborative latent structure learning and collaborative relationship reconstruction process.

[0120] During the training of the improved VGAE, multiple surface image samples of injection-molded products are collected to generate corresponding defect structure diagrams and defect action mode sets. The graph nodes and edges in the defect structure diagrams and the defect action modes in the defect action mode sets are organized according to their correspondences to form a model training sample set. Based on the model training sample set, an action mode node set and an action mode connection structure are established. The action mode node set represents multiple defect action modes, and the action mode connection structure represents the mode connection relationships between different action mode nodes. An improved VGAE is constructed based on the action mode node set and the action mode connection structure. The improved VGAE includes an action mode implicit structure generation layer and an action mode collaborative reconstruction layer. The action mode node set and the action mode connection structure are input into the action mode implicit structure generation layer. The changing behavior patterns and propagation behavior patterns corresponding to the action mode nodes are used as learning content, and the mode connection relationships are used as information propagation paths to generate corresponding... The latent space representation results of the action pattern nodes and potential collaborative relationships are used to form a collaborative latent structure representation of the action pattern. The collaborative latent structure representation of the action pattern is input into the action pattern collaborative reconstruction layer, and in combination with the collaborative reconstruction rule set, collaborative recovery processing is performed on the collaborative relationships between action pattern nodes to generate corresponding collaborative reconstruction results. A reconstruction loss function is constructed based on the structural reconstruction error between the collaborative reconstruction results and the original action pattern connection structure, and a collaborative loss function is constructed based on the collaborative consistency error between the potential collaborative relationships and the pattern connection relationship. A total loss function is constructed based on the reconstruction loss function and the collaborative loss function. The Adam optimizer is used to iteratively update the improved VGAE parameters, and the validation loss is calculated after each training epoch. When the change of the validation loss in multiple consecutive training epochs is lower than the preset convergence condition or the training epochs reach the training limit, the model training is stopped, the model parameters corresponding to the minimum validation loss are saved, and the improved VGAE with complete training is obtained.

[0121] In this embodiment, the generation of the cooperative implicit structure of the action mode specifically includes:

[0122] Map multiple graph nodes in the defect structure graph to multiple defect action modes in the defect action mode set to generate action mode mapping results;

[0123] Based on the effect mode mapping results, the graph nodes corresponding to the same defect effect mode are merged to form multiple effect mode nodes, and the mode connection relationship between the multiple effect mode nodes is established.

[0124] Input multiple action mode nodes and mode connection relationships into the action mode implicit structure generation layer in the improved VGAE to generate the action mode propagation structure.

[0125] The implicit structure generation layer of the action pattern is used to perform implicit space propagation processing on multiple action pattern nodes in the action pattern propagation structure to generate action pattern collaborative representation results.

[0126] The generation of the collaborative characterization results of the action mode specifically includes:

[0127] Read multiple action mode nodes and their connection relationships in the action mode propagation structure; determine the connection status between different action mode nodes based on the connection relationships; generate action mode collaborative representation results based on the connection status between different action mode nodes; the action mode collaborative representation results are used to record the connection status between different action mode nodes.

[0128] Based on the collaborative characterization results of action modes, the latent space relationships between nodes of different action modes are extracted to generate the action mode association structure.

[0129] The generation of the action mode association structure specifically includes:

[0130] Read multiple action pattern nodes and their corresponding connections from the action pattern collaboration representation results; identify action pattern nodes with common connections based on the connections between different action pattern nodes; determine action pattern nodes with common connections as action pattern nodes with latent space associations; perform unified organization processing on multiple latent space associations to generate an action pattern association structure; latent space associations are used to represent potential collaboration relationships between different action pattern nodes; the action pattern association structure is used to record the latent space association relationships between multiple action pattern nodes.

[0131] Perform structural organization processing on multiple implicit space relationships in the action mode association structure to generate an action mode cooperative implicit structure.

[0132] The generation of cooperative implicit structures based on action modes specifically includes:

[0133] Read multiple latent space relationships in the action mode association structure; determine the corresponding action mode nodes based on each latent space relationship; organize the corresponding action mode nodes and corresponding latent space relationships in a unified manner, so that each latent space relationship and the corresponding action mode node form a corresponding association structure; generate the action mode cooperative latent structure based on the formed corresponding association structure.

[0134] The action pattern association structure is used to record the latent space association relationships between different action pattern nodes; the action pattern cooperative latent structure is used to record the overall structure formed by multiple action pattern nodes and their corresponding latent space association relationships; since the action pattern association structure only records latent space association relationships and cannot fully represent the action pattern nodes associated with the latent space association relationships, the action pattern nodes and their corresponding latent space association relationships are organized in a unified manner to generate the action pattern cooperative latent structure.

[0135] In this embodiment, the generation of the collaborative reconstruction graph of the action mode specifically includes:

[0136] Read multiple action pattern nodes and multiple latent space relationships in the collaborative implicit structure of action patterns, and generate collaborative reconstruction input results;

[0137] The collaborative reconstruction input results are input into the action mode collaborative reconstruction layer, and collaborative reconstruction processing is performed on the latent space association relationships between multiple action mode nodes to generate action mode collaborative reconstruction results.

[0138] The generation of the collaborative reconstruction results of the action mode specifically includes:

[0139] Read multiple action pattern nodes and multiple latent space relationships from the collaborative reconstruction input results; determine the corresponding connected action pattern nodes based on each latent space relationship; count the association of each action pattern node in different latent space relationships and generate node association results; identify action pattern nodes with common latent space relationships around each node association result and generate node collaboration results; perform unified organization processing on multiple node collaboration results to generate action pattern collaborative reconstruction results.

[0140] The collaborative reconstruction results of action modes are used to record the collaborative situation between nodes of different action modes. The collaborative reconstruction results of action modes include action mode nodes with collaborative associations and corresponding latent space associations. Each action mode node is used to characterize the corresponding defect action mode, and each latent space association is used to characterize the latent space collaborative state between the corresponding action mode nodes.

[0141] Based on the collaborative reconstruction results of the action mode, the collaborative association state between multiple action mode nodes is restored, and the collaborative association structure of the action mode is generated.

[0142] The generation of the collaborative association structure of the action mode specifically includes:

[0143] Read the action pattern nodes and their corresponding latent space relationships from the action pattern collaborative reconstruction results, and restore the collaborative relationship state between different action pattern nodes based on the latent space relationship; then generate the action pattern collaborative relationship structure based on multiple action pattern nodes and their corresponding collaborative relationship states.

[0144] The action mode collaborative association structure is used to characterize the collaborative association state between nodes with different action modes;

[0145] Perform graph structure mapping processing on the collaborative association structure of action modes to generate a collaborative reconstruction graph of action modes;

[0146] The generation of the collaborative reconstruction graph of action modes specifically includes:

[0147] Read multiple action mode nodes and their corresponding collaborative association states from the action mode collaborative association structure; map the multiple action mode nodes to corresponding graph nodes, and map the multiple collaborative association states to corresponding graph edges; generate an action mode collaborative reconstruction graph based on the graph nodes and graph edges; the action mode collaborative reconstruction graph is used to represent the collaborative association between different action mode nodes.

[0148] In this embodiment, the generation of the abnormal action mode cooperative structure specifically includes:

[0149] Read multiple reconstructed graph nodes and edges from the collaborative reconstructed graph of the action mode, and read multiple graph nodes and edges from the defect structure graph to establish the correspondence between the collaborative reconstructed graph of the action mode and the defect structure graph, and generate the graph structure correspondence result.

[0150] Based on the graph structure correspondence results, node difference analysis is performed on multiple reconstructed graph nodes and their corresponding graph nodes to extract the feature deviation states between the corresponding nodes and generate node difference results;

[0151] The generation of node difference results specifically includes:

[0152] Read multiple reconstructed graph nodes and their corresponding graph nodes from the graph structure corresponding results; extract the spatial location information, texture feature information, edge feature information, and structural feature information of each reconstructed graph node and its corresponding graph node; calculate the positional offset between corresponding spatial location information, the texture difference between corresponding texture feature information, the edge difference between corresponding edge feature information, and the structural difference between corresponding structural feature information; generate the feature deviation state between corresponding nodes based on the positional offset, texture difference, edge difference, and structural difference; perform unified recording processing on multiple feature deviation states to generate node difference results;

[0153] Perform edge association difference analysis on multiple reconstructed graph edges and their corresponding graph edges, extract the association deviation state between corresponding graph edges, and generate edge difference results;

[0154] The generation of edge difference results specifically includes:

[0155] Read multiple reconstructed graph edges and their corresponding edges from the graph structure corresponding results; extract the spatial adjacency, texture similarity, edge continuity, and structural association relationships between each reconstructed graph edge and its corresponding edge; calculate the adjacency difference between corresponding spatial adjacency relationships, the texture association difference between corresponding texture similarity relationships, the edge association difference between corresponding edge continuity relationships, and the structural association difference between corresponding structural association relationships; generate the association deviation state between corresponding graph edges based on the adjacency difference, texture association difference, edge association difference, and structural association difference; perform unified recording processing on multiple association deviation states to generate edge difference results; the edge difference results are used to record the association deviation between each reconstructed graph edge and its corresponding edge.

[0156] The node difference results and edge difference results are fused together to generate a collaborative difference result;

[0157] The generation of collaborative discrepancy results specifically includes:

[0158] Read multiple feature deviation states from the node difference results and multiple associated deviation states from the edge difference results; determine the associated deviation states corresponding to each feature deviation state based on the graph structure correspondence results; perform corresponding association processing on the feature deviation states corresponding to the same reconstructed graph node and the associated deviation states corresponding to its associated reconstructed graph edges to generate the corresponding node's collaborative deviation states; perform unified recording processing on multiple collaborative deviation states to generate collaborative difference results; the collaborative difference results are used to record the joint deviation of reconstructed graph nodes and their associated reconstructed graph edges; the collaborative difference results include the feature deviation states of the corresponding node and the associated deviation states of the graph edges associated with that node;

[0159] Based on the results of collaborative differences, the reconstructed graph nodes with collaborative deviation states and their corresponding reconstructed graph edges are located, and a set of nodes with abnormal action patterns and a set of edges associated with abnormal action patterns are generated.

[0160] The generation of the set of nodes representing abnormal action patterns and the set of edges associated with abnormal action patterns specifically includes:

[0161] Read the collaborative difference results; locate the reconstructed graph nodes with node and edge deviation based on the collaborative difference results; organize the located reconstructed graph nodes into an abnormal action mode node set; extract the reconstructed graph edges associated with the abnormal action mode node set and organize them into an abnormal action mode associated edge set.

[0162] Based on the set of nodes and edges associated with the abnormal action patterns, structural organization processing is performed to generate a collaborative structure for the abnormal action patterns.

[0163] The generation of the abnormal action mode cooperative structure specifically includes:

[0164] Read the set of nodes and edges associated with the abnormal action pattern; organize the nodes according to the connection relationship of the corresponding edges; construct a complete cooperative structure of the abnormal action pattern based on the correspondence between nodes and edges;

[0165] The abnormal action pattern collaborative structure is used to characterize the distribution of abnormal action patterns corresponding to surface defects of injection molded products. The abnormal action pattern node set can only record action pattern nodes with abnormal deviations, and the abnormal action pattern associated edge set can only record the association between corresponding action pattern nodes, but cannot fully characterize the overall collaborative abnormal state between multiple abnormal action pattern nodes. Therefore, the abnormal action pattern node set and the abnormal action pattern associated edge set are organized in a unified manner to generate the abnormal action pattern collaborative structure. The abnormal action pattern collaborative structure is used to record the collaborative association between abnormal action pattern nodes and the abnormal propagation, serving as the data basis for subsequently generating the surface image defect recognition results of injection molded products.

[0166] In this embodiment, the generation of surface image defect recognition results for injection molded products specifically includes:

[0167] Read multiple abnormal action mode nodes and multiple abnormal action mode associated edges in the abnormal action mode collaborative structure, establish the mapping relationship between abnormal action mode nodes and corresponding defect candidate regions, and generate abnormal region results.

[0168] Based on the abnormal region results, extract the texture change features, edge change features, and structural change features from the corresponding defect candidate regions to generate abnormal feature results;

[0169] Based on the abnormal feature results and the collaborative deviation state corresponding to the associated edge of the abnormal action mode, defect category determination processing is performed to generate defect category results;

[0170] The defect location result is generated by combining the defect category result with the location distribution status of the corresponding defect candidate region.

[0171] Based on the defect category and defect location results, generate defect recognition results for the surface image of the injection molded product.

[0172] Example 1: To verify the feasibility of this invention in practice, it was applied to the surface quality inspection of injection molded products in an injection molding workshop in East China. This workshop produces electronic product casings, household appliance structural components, and industrial plastic assemblies. Due to the influence of factors such as mold condition, raw material flow, cooling conditions, and fluctuations in molding process parameters during injection molding, defects such as shrinkage marks, cracks, bubbles, flow marks, deformation, and localized depressions are easily generated on the product surface. Since different defects often do not exist independently during their formation but rather change and influence each other along with adjacent areas, multiple defect areas usually have a certain propagation correlation and co-evolution relationship. Existing detection methods mostly rely on manual observation or identification based on local image features. While these methods can identify defects with obvious features, they are prone to omissions and misjudgments when the defect area is small, the features are weak, or multiple defects occur simultaneously, making it difficult to accurately reflect the inherent correlation between defects.

[0173] In practical applications, industrial image acquisition equipment is installed in the inspection area of ​​the production workshop to acquire images of the surface of injection-molded products passing through the inspection station during continuous production. This yields surface image data containing surface texture, edge, and grayscale information. The acquisition time covers multiple production shifts, and the acquisition location is situated in the quality inspection area after injection molding, ensuring that the acquired images accurately reflect the surface condition of the products. Subsequently, this invention is used to preprocess the acquired surface image data, performing image denoising, size standardization, brightness correction, and image enhancement to generate a standardized surface image dataset. Analysis of the standardized surface image dataset identifies texture anomaly regions, edge anomaly regions, and structural anomaly regions, and these anomaly regions are correlated to form a set of defect candidate regions.

[0174] After obtaining the set of candidate defect regions, this invention further extracts the spatial location information, texture feature information, edge feature information, and structural feature information corresponding to each candidate defect region. Based on the proximity, texture variation, edge extension, and structural association between different candidate defect regions, a correspondence is established to construct a defect structure graph. At this point, different candidate defect regions are no longer analyzed as isolated regions, but are organized into a structural network with overall relationships, thus reflecting the association state between multiple defect regions. Subsequently, based on the graph node features and graph edge association features in the defect structure graph, action feature vectors are constructed. The consistency of change and the cooperative propagation state between different action feature vectors are further analyzed. Action feature vectors with the same change behavior and propagation behavior are aggregated to generate a corresponding set of defect action patterns.

[0175] After the defect action pattern set is generated, the defect structure diagram and the defect action pattern set are input into the improved VGAE for processing. Unlike traditional methods that directly analyze the connections between defect regions, this invention first establishes a set of action pattern nodes and an action pattern connection structure. Through an action pattern hidden structure generation layer, it learns the cooperative change relationships and propagation associations between different action patterns, forming a cooperative hidden structure of action patterns. Based on this, a cooperative reconstruction layer of action patterns is used to restore the cooperative association states between different action pattern nodes, generating a cooperative reconstruction graph of action patterns. Through this process, not only can the revealed defect associations be identified, but also the potential cooperative relationships hidden between multiple defect regions can be discovered, thereby achieving a deep characterization of the formation mechanism of complex defects.

[0176] Subsequently, a corresponding analysis was performed on the collaborative reconstruction graph of the action mode and the defect structure graph, extracting node difference results and edge difference results respectively, and further merging them to form a collaborative difference result. Based on the collaborative difference result, abnormal action mode nodes and associated edges were located, ultimately constructing an abnormal action mode collaborative structure. Since the abnormal action mode collaborative structure simultaneously records the association between abnormal nodes and the abnormal propagation, it can more completely reflect the propagation characteristics and collaborative change characteristics of defects in different regions. Based on the abnormal region information and abnormal feature information in the abnormal action mode collaborative structure, the texture change characteristics, edge change characteristics, and structural change characteristics corresponding to different defect candidate regions were comprehensively judged, thereby generating corresponding defect category results and defect location results, ultimately forming the defect recognition result for the injection molded product surface image.

[0177] During continuous operation, surface images of injection-molded products acquired across multiple production time periods were processed, and the identification results were compared with quality inspection records. Results show that this invention can stably identify multiple different types of surface defects and effectively detect collaborative anomaly relationships between multiple defect regions. When the local anomaly features exhibited by some defect regions are not obvious, this invention can still identify potential anomaly regions by utilizing the collaborative correlation between action modes, thus avoiding missed detections caused by relying solely on local feature analysis. Furthermore, for situations where multiple defects coexist and influence each other, this invention can reveal the propagation correlation state between different defects through the collaborative hidden structure of action modes and the collaborative reconstruction diagram of action modes, improving the identification capability in complex defect scenarios.

[0178] Through application in actual production environments, this invention has demonstrated its ability to stably identify surface defects generated across different time periods, batches, and process conditions, while maintaining consistent results in defect correlation analysis. By introducing a defect action mode set, a collaborative implicit structure of action modes, and a collaborative reconstruction mechanism of action modes, the defect identification process not only focuses on local regional anomalies but also analyzes the formation and propagation patterns of defects from an overall collaborative perspective, thus providing a more complete characterization of the defect evolution process. Practical application data shows that the defect identification results obtained by this invention exhibit good consistency with on-site quality inspection results, effectively improving the accuracy of identification, anomaly pattern detection capability, and stability of identification results in complex defect scenarios. This verifies the practical application value and feasibility of this invention in the field of surface image defect identification for injection molded products.

[0179] Table 1. Performance comparison of different defect identification methods in the scenario of surface defect detection of injection molded products.

[0180] Evaluation indicators Texture feature-based methods CNN-based methods Graph structure analysis method Method of the present invention Defect identification accuracy (%) 72.8 79.6 83.7 88.4 Defect recall rate (%) 69.5 76.2 81.1 86.7 Defect category identification consistency rate (%) 71.3 78.4 82.6 87.9 Weak feature defect identification rate (%) 58.7 67.9 74.5 82.3 Multi-defect scene recognition rate (%) 61.8 70.6 77.2 85.1 Defect area location overlap rate (%) 66.9 74.8 80.5 86.2 Anomaly association detection rate (%) 49.6 58.3 71.4 84.7 Stability coefficient of recognition results 0.63 0.71 0.78 0.86 Average processing time per image (s) 1.92 1.48 1.31 1.09

[0181] As shown in Table 1, the method of this invention exhibits significant advantages across all evaluation metrics. Specifically, in terms of defect recognition accuracy, this invention achieves 88.4%, which is 15.6 percentage points higher than texture-based methods, 8.8 percentage points higher than CNN-based methods, and 4.7 percentage points higher than graph structure analysis-based methods. This demonstrates that this invention not only utilizes the texture, edge, and structural features of the defect region itself but also further explores the collaborative relationships between different defect regions, thus obtaining more complete defect characterization information and improving the final recognition accuracy.

[0182] In terms of defect recall rate and defect category identification consistency rate, the present invention achieves 86.7% and 87.9% respectively. Compared with traditional methods, the present invention has a stronger coverage of defect samples, especially in the case of complex defects and defects with ambiguous boundaries. It can avoid some defects being missed. The reason is that the present invention constructs a defect structure diagram and a set of defect action modes. It is no longer limited to local analysis of a single region, but analyzes the defect formation process from the perspective of overall correlation. Therefore, it has a better ability to discover hidden defects and coupled defects.

[0183] In terms of weak feature defect recognition rate, the present invention achieves 82.3%, which is 23.6 percentage points higher than the traditional texture feature method. Weak feature defects such as slight shrinkage marks, shallow flow marks and early cracks often occur in the injection molding process. These defects are not obvious in the local manifestation of the image, and it is difficult to accurately identify them by relying on single region features. The present invention learns the potential synergistic relationship between different action modes through the action mode hidden structure generation layer, so that the originally inconspicuous abnormal areas can be further strengthened through the correlation propagation relationship, thus significantly improving the recognition ability of weak feature defects.

[0184] In terms of multi-defect scene recognition rate, the present invention achieves 85.1%, which is 7.9 percentage points higher than the graph structure analysis method. In actual production, multiple defects often coexist and there are mutual influence relationships between different defects. Traditional methods can usually only identify a single defect area and it is difficult to analyze the co-evolution state between defects. The present invention utilizes the action mode co-cooperative hidden structure and the action mode co-reconstruction graph to restore the co-cooperative relationship between different action modes, so that multiple defects can be analyzed in the form of an overall structure. Therefore, it has better recognition ability in multi-defect scenes.

[0185] The technical advantages of this invention can be further demonstrated by the overlap rate of defect region localization and the anomaly correlation detection rate. This invention achieves a defect region localization overlap rate of 86.2% and an anomaly correlation detection rate of 84.7%, showing a significant improvement compared to traditional methods. Specifically, the anomaly correlation detection rate is 13.3 percentage points higher than methods based on graph structure analysis. This indicates that this invention can not only detect the defects themselves but also uncover the hidden propagation correlation states between multiple defect regions. The fundamental reason lies in the introduction of a collaborative reconstruction mechanism of action modes in this invention. By restoring explicit and implicit collaborative relationships, the collaborative anomaly structure hidden behind the defects is revealed.

[0186] Furthermore, the recognition results of this invention achieve a stability coefficient of 0.86, maintaining good recognition consistency across images acquired at different time periods. Simultaneously, the average processing time per image is only 1.09 seconds, approximately 43.2% shorter than texture feature-based methods and approximately 26.4% shorter than CNN-based methods. This demonstrates that this invention improves recognition accuracy without significantly increasing computational burden. Because the action pattern set aggregates and represents a large number of similar defect behaviors, the network can perform collaborative analysis at a higher level, reducing redundant computation and thus balancing recognition efficiency and accuracy.

[0187] As can be seen from the data in Table 1, this invention, by constructing a defect structure diagram, a defect action mode set, and an improved VGAE containing an action mode hidden structure generation layer and an action mode collaborative reconstruction layer, achieves a transformation from "defect region analysis" to "defect action mode collaborative analysis." This not only improves the accuracy of defect identification, the ability to identify weak feature defects, and the adaptability to multiple defect scenarios, but also enhances the ability to detect anomalies and the stability of identification. This fully demonstrates the technological advancement and practical application value of this invention in the field of surface image defect recognition of injection molded products.

[0188] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying surface image defects in injection-molded products based on deep learning, characterized in that, Includes the following steps: Collect surface image data of injection molded products and preprocess them to generate a standardized surface image dataset; Perform defect candidate region extraction processing on a standardized surface image dataset to generate a set of defect candidate regions; A defect structure graph is constructed based on the relationship between multiple defect candidate regions. The defect candidate regions are mapped as graph nodes, and the relationship between the candidate regions is mapped as graph edges to generate the defect structure graph. Based on the graph node features and graph edge association features in the defect structure graph, an action feature vector is constructed. Action behavior representation processing is performed on multiple action feature vectors to generate a set of defect action patterns. An improved VGAE is constructed, which includes an action mode implicit structure generation layer and an action mode cooperative reconstruction layer. The defect structure diagram and the defect action mode set are input into the improved VGAE, and the action mode cooperative implicit structure is generated through the action mode implicit structure generation layer. The collaborative reconstruction layer of action mode performs collaborative reconstruction processing based on the implicit structure of action mode collaboration, and generates an action mode collaborative reconstruction graph. Perform a collaborative difference analysis on the collaborative reconstruction diagram of action modes and the defect structure diagram to generate a collaborative structure of abnormal action modes; Based on the abnormal action mode, the surface image defect recognition results of injection molded products are generated.

2. The method for identifying surface image defects of injection molded products based on deep learning according to claim 1, characterized in that, The surface image data specifically includes surface texture images, surface edge images, and surface grayscale images; the preprocessing specifically includes image denoising, image size unification, image brightness correction, and image enhancement.

3. The method for identifying surface image defects of injection molded products based on deep learning according to claim 1, characterized in that, The generation of the defect candidate region set specifically includes: Texture anomaly region extraction is performed on surface texture images in the standardized surface image dataset to generate texture anomaly regions; edge anomaly region extraction is performed on surface edge images to generate edge anomaly regions; structural anomaly region extraction is performed on surface grayscale images to generate structural anomaly regions; region correspondence analysis is performed on texture anomaly regions, edge anomaly regions, and structural anomaly regions to generate candidate defect regions; and region organization processing is performed on multiple candidate defect regions to generate a set of candidate defect regions.

4. The method for identifying surface image defects of injection molded products based on deep learning according to claim 1, characterized in that, The generation of the defect structure diagram specifically includes: Extract multiple candidate defect regions from the defect candidate region set, and determine the spatial location information, texture feature information, edge feature information, and structural feature information corresponding to each candidate defect region; generate candidate association relationships based on the spatial adjacency relationship, texture similarity relationship, edge continuity relationship, and structural association relationship between multiple candidate defect regions; map multiple candidate defect regions to corresponding graph nodes, and map multiple candidate association relationships to corresponding graph edges to generate a defect structure graph.

5. The method for identifying surface image defects of injection molded products based on deep learning according to claim 1, characterized in that, The generation of the defect action mode set specifically includes: Read multiple graph nodes and multiple graph edges in the defect structure graph, extract the graph node features corresponding to each graph node and the graph edge association features corresponding to each graph edge, and generate a feature data set. Based on the graph node features and graph edge association features in the feature data set, construct corresponding action feature vectors and generate an action feature vector set; Perform change consistency analysis and propagation synergy analysis on multiple action feature vectors in the action feature vector set to generate action behavior representation results; Based on the behavioral representation results, behavioral aggregation is performed on the behavioral feature vectors with the same behavioral characteristics to generate multiple defect behavioral patterns; Perform pattern organization processing on multiple defect action patterns to generate a defect action pattern set.

6. The method for identifying surface image defects of injection molded products based on deep learning according to claim 1, characterized in that, The construction of the improved VGAE specifically includes: Extract multiple defect action patterns from the defect action pattern set, establish action pattern nodes based on the correspondence between the action behaviors of multiple defect action patterns, and generate an action pattern node set. Around multiple action pattern nodes in the action pattern node set, extract the collaborative change state and propagation association state between different action pattern nodes, establish the pattern connection relationship between action pattern nodes, and generate the action pattern connection structure. Based on the set of action pattern nodes and the action pattern connection structure, an action pattern hidden structure generation layer is constructed. The action pattern nodes are used as the hidden structure learning objects, and the pattern connection relationship is used as the hidden structure propagation path to generate the action pattern hidden structure generation layer. Read the set of action mode nodes and the action mode connection structure, extract the collaborative association state and collaborative propagation state between each action mode node, construct action mode collaborative reconstruction rules, and generate a set of collaborative reconstruction rules. Based on the set of action mode nodes, the action mode connection structure, and the set of collaborative reconstruction rules, an action mode collaborative reconstruction layer is constructed, and the collaborative relationship between action mode nodes is used as the reconstruction object to generate the action mode collaborative reconstruction layer. Establish the data transfer relationship between the implicit structure generation layer of the action mode and the collaborative reconstruction layer of the action mode, integrate the implicit structure generation layer of the action mode and the collaborative reconstruction layer of the action mode, and generate an improved VGAE.

7. The method for identifying surface image defects of injection molded products based on deep learning according to claim 1, characterized in that, The generation of the cooperative implicit structure of the aforementioned action mode specifically includes: Map multiple graph nodes in the defect structure graph to multiple defect action modes in the defect action mode set to generate action mode mapping results; Based on the effect mode mapping results, the graph nodes corresponding to the same defect effect mode are merged to form multiple effect mode nodes, and the mode connection relationship between the multiple effect mode nodes is established. Input multiple action mode nodes and mode connection relationships into the action mode implicit structure generation layer in the improved VGAE to generate the action mode propagation structure. The implicit structure generation layer of the action pattern is used to perform implicit space propagation processing on multiple action pattern nodes in the action pattern propagation structure to generate action pattern collaborative representation results. Based on the collaborative characterization results of action modes, the latent space relationships between nodes of different action modes are extracted to generate the action mode association structure. Perform structural organization processing on multiple implicit space relationships in the action mode association structure to generate an action mode cooperative implicit structure.

8. The method for identifying surface image defects of injection molded products based on deep learning according to claim 1, characterized in that, The generation of the collaborative reconstruction graph of the action mode specifically includes: Read multiple action pattern nodes and multiple latent space relationships in the collaborative implicit structure of action patterns, and generate collaborative reconstruction input results; The collaborative reconstruction input results are input into the action mode collaborative reconstruction layer, and collaborative reconstruction processing is performed on the latent space association relationships between multiple action mode nodes to generate action mode collaborative reconstruction results. Based on the collaborative reconstruction results of the action mode, the collaborative association state between multiple action mode nodes is restored, and the collaborative association structure of the action mode is generated. Perform graph structure mapping processing on the action mode collaborative association structure to generate an action mode collaborative reconstruction graph.

9. The method for identifying surface image defects of injection molded products based on deep learning according to claim 1, characterized in that, The generation of the abnormal action mode cooperative structure specifically includes: Read multiple reconstructed graph nodes and edges from the collaborative reconstructed graph of the action mode, and read multiple graph nodes and edges from the defect structure graph to establish the correspondence between the collaborative reconstructed graph of the action mode and the defect structure graph, and generate the graph structure correspondence result. Based on the graph structure correspondence results, node difference analysis is performed on multiple reconstructed graph nodes and their corresponding graph nodes to extract the feature deviation states between the corresponding nodes and generate node difference results; Perform edge association difference analysis on multiple reconstructed graph edges and their corresponding graph edges, extract the association deviation state between corresponding graph edges, and generate edge difference results; The node difference results and edge difference results are fused together to generate a collaborative difference result; Based on the results of collaborative differences, the reconstructed graph nodes with collaborative deviation states and their corresponding reconstructed graph edges are located, and a set of nodes with abnormal action patterns and a set of edges associated with abnormal action patterns are generated. Based on the set of nodes and edges associated with the abnormal action patterns, structural organization processing is performed to generate a collaborative structure for the abnormal action patterns.

10. The method for identifying surface image defects of injection molded products based on deep learning according to claim 1, characterized in that, The generation of the surface image defect recognition results of the injection molded product specifically includes: Read multiple abnormal action mode nodes and multiple abnormal action mode associated edges in the abnormal action mode collaborative structure, establish the mapping relationship between abnormal action mode nodes and corresponding defect candidate regions, and generate abnormal region results. Based on the abnormal region results, extract the texture change features, edge change features, and structural change features from the corresponding defect candidate regions to generate abnormal feature results; Based on the abnormal feature results and the collaborative deviation state corresponding to the associated edge of the abnormal action mode, defect category determination processing is performed to generate defect category results; The defect location result is generated by combining the defect category result with the location distribution status of the corresponding defect candidate region. Based on the defect category and defect location results, generate defect recognition results for the surface image of the injection molded product.

Images