Picture content compliance review method and system based on ai inference

Abstract

The present application relates to the technical field of picture content review, and more particularly to a picture content compliance review method and system based on AI inference, which generates context-aware scene semantic representation through multi-granularity segmentation and relationship modeling, performs explicit and implicit risk detection in parallel, and uses a dynamic rule knowledge graph for reasoning and conflict identification to generate graded decisions and causal explanation chains. The method can improve the accuracy and explainability of the review, and realize dynamic optimization of the review rules.

Description

Technical Field

[0001] This invention relates to the field of image content review technology, and in particular to an image content compliance review method and system based on AI reasoning. Background Technology

[0002] In the field of internet content management, the compliance review of image content is a crucial link in ensuring the security and order of the online environment. Currently, the mainstream review methods mainly rely on deep learning-based computer vision models, which are trained on large-scale labeled datasets to identify illegal content in images, such as violence, pornography, or sensitive symbols.

[0003] These conventional methods typically employ an end-to-end classification or detection framework, inputting the entire image or a predefined region into a neural network and directly outputting the violation probability or violation category label. Their technical implementation focuses on learning visual features at the pixel level and relies on the model's performance on a closed test set to evaluate the review effectiveness. To handle complex scenarios, some solutions introduce object detection techniques, first identifying key entities in the image and then classifying these entities separately.

[0004] However, existing practices have significant limitations. On the one hand, their review logic heavily relies on surface visual pattern associations learned by the model from training data, lacking an understanding of the deep semantics and contextual information of images. The model struggles to resolve complex relationships between objects; for example, it might identify "knife" and "person" as isolated objects and classify them as posing a risk of violence, but fail to distinguish between a kitchen cooking scene and a violent threat scene, leading to rigid and mechanical review results with a high misjudgment rate. This mechanism, based on statistical association rather than causal reasoning, makes the review process a "black box," unable to provide clear and traceable explanatory evidence for review decisions.

[0005] On the other hand, the rules and models of existing systems are usually static and isolated. Review rules often exist in the form of hard-coded thresholds or keyword lists, which are out of sync with the update cycles of the AI ​​models responsible for feature extraction, making it difficult to adapt to the ever-evolving new violations and semantic boundaries in the online environment. When "gray areas" not covered by the rules appear or when potential conflicts exist between different rules, the system lacks dynamic reasoning and coordination mechanisms, often relying solely on manual review for case-by-case processing. This results in limited automation and intelligence, making it difficult for the system as a whole to achieve continuous adaptive optimization. Summary of the Invention

[0006] This invention provides a method and system for reviewing image content compliance based on AI reasoning, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a method for verifying the compliance of image content based on AI inference, comprising: The process involves acquiring images to be reviewed and their associated multi-source context data, performing multi-granularity region segmentation and object relationship modeling on the images to be reviewed, extracting spatial topological relationships and semantic dependencies between object entities in the images, generating a structured scene graph representation, and applying context-sensitive weighting to the nodes in the structured scene graph representation based on the multi-source context data to form a context-aware scene semantic representation. The scene semantic representation is input into the reasoning module, and explicit violation detection and implicit risk prediction are performed in parallel based on the spatial topological relationship and semantic dependency relationship in the scene semantic representation to generate multi-dimensional compliance assessment results. A dynamic rule knowledge graph is established. Based on the multi-dimensional compliance assessment results, reasoning path search is performed in the dynamic rule knowledge graph to identify the triggered rule chains and conflict relationships between rules, generate hierarchical review decisions, and extract the rule nodes in the reasoning path and the corresponding evidence nodes in the scene semantic representation to construct a causal explanation chain for the review decision. The hierarchical review decision is executed, and manual review feedback and user appeal data are collected. The manual review feedback and user appeal data are used to incrementally update the rule entity relationships in the dynamic rule knowledge graph. At the same time, the context sensitivity weight in the scene semantic representation is adjusted backtrackingly according to the causal explanation chain.

[0008] The images to be reviewed are subjected to multi-granularity region segmentation and object relationship modeling. Spatial topological relationships and semantic dependencies between objects in the images are extracted, and a structured scene graph representation is generated, including: A multi-scale feature pyramid is constructed, and the image to be reviewed is segmented into hierarchical regions using the multi-scale feature pyramid. Candidate region sets are generated at different scale levels, and cross-level consistency checks are performed on the candidate regions at each scale level to select a stable region set with scale invariance. Object entity identification and attribute labeling are performed on each region in the stable region set to obtain an object entity set containing object category, location coordinates and semantic attributes. The relative orientation and spatial distance between objects are calculated based on the location coordinates of each object entity in the object entity set to determine the spatial topological relationship. The semantic attributes of each object entity in the object entity set are mapped to a pre-constructed semantic knowledge space. The semantic distance and co-occurrence probability between object entities in the semantic knowledge space are calculated. The semantic distance is then spatially constrained and corrected in conjunction with the spatial topological relationship to generate a relation weight matrix representing semantic dependencies. A structured scene graph representation is constructed using the object entities in the object entity set as nodes and the spatial topological relationship and the relationship weight matrix as the attributes of the edges.

[0009] A multi-scale feature pyramid is constructed, and the image to be reviewed is segmented into hierarchical regions using the multi-scale feature pyramid. Candidate region sets are generated at different scale levels, and cross-level consistency checks are performed on the candidate regions at each scale level. The set of stable regions with scale invariance is selected, including: The images to be reviewed are downsampled step by step to generate multiple image layers with decreasing resolution. Semantic features and structural features are extracted from each image layer. A bidirectional feature transfer path is established between adjacent image layers. The semantic features and structural features of different image layers are fused across layers through the bidirectional feature transfer path to construct a multi-scale feature pyramid. Region extraction operations are performed at each image level of the multi-scale feature pyramid to generate a candidate region set corresponding to each image level. Cross-level spatial alignment is performed on the candidate region set, projecting candidate regions from different image levels onto a unified reference coordinate system. Cross-level candidate region pairs with overlapping spatial positions in the unified reference coordinate system are identified. The feature response consistency of each cross-level candidate region pair is calculated. When the feature response consistency meets a preset stability condition, the corresponding cross-level candidate region pairs are merged into a single stable region. The single stable regions are then summarized to form a stable region set.

[0010] Based on the multi-source context data, the nodes in the structured scene graph representation are weighted by context sensitivity to form a context-aware scene semantic representation, including: Time series features, geographic location features, and user interaction features are extracted from multi-source context data, and cross-modal semantic alignment is performed on the time series features, geographic location features, and user interaction features to generate a fused context semantic embedding vector. The semantic representation of each node in the structured scene graph representation is matched with the context semantic embedding vector for similarity, and the semantic activation strength of each node under the current context condition is calculated. Based on the semantic activation intensity, a node importance ranking sequence is constructed, and a node sensitivity weight is calculated according to the relative position of each node in the node importance ranking sequence. The node sensitivity weight is then applied to the feature representation of the corresponding node in the structured scene graph representation to generate a context-weighted node feature representation. The node feature representations and their associated edge relationships are aggregated by graph convolution to form a context-aware scene semantic representation.

[0011] The scene semantic representation is input into the inference module. Based on the spatial topological relationships and semantic dependencies in the scene semantic representation, explicit violation detection and implicit risk prediction are performed in parallel to generate multi-dimensional compliance assessment results, including: Extract object location distribution features corresponding to spatial topological relationships and object interaction pattern features corresponding to semantic dependencies from scene semantic representation, and input the object location distribution features into the explicit violation detection branch and the object interaction pattern features into the implicit risk prediction branch. In the explicit violation detection branch, the location distribution features of the object are matched with rules based on a pre-built violation rule base to identify explicit violation objects that violate explicit prohibitions and generate explicit violation detection results; In the implicit risk prediction branch, a behavioral association graph between objects is constructed based on the object interaction pattern characteristics. By analyzing the abnormal association patterns in the behavioral association graph, objects with potential hidden risks that could lead to violations are identified, and implicit risk prediction results are generated. The explicit violation detection results are integrated with the implicit risk prediction results, and the explicit violation objects and the implicit risk objects are respectively assigned violation severity scores and risk tendency scores to generate a multi-dimensional compliance assessment result that includes the violation severity scores and the risk tendency scores.

[0012] A dynamic rule knowledge graph is established. Based on the multi-dimensional compliance assessment results, reasoning path searches are performed within the dynamic rule knowledge graph to identify triggered rule chains and conflict relationships between rules, generating tiered review decisions, including: A dynamic rule knowledge graph containing rule nodes and rule relationship edges is constructed, wherein the rule relationship edges represent the logical dependencies and temporal triggering relationships between rules, and the dynamic rule knowledge graph is updated in real time to reflect the dynamic changes of the rule base. Using the severity score and risk propensity score from the multi-dimensional compliance assessment results as query conditions, the initial set of rule nodes that match the severity score and risk propensity score are located in the dynamic rule knowledge graph. Starting from the initial set of rule nodes, a graph traversal expansion is performed along the rule relationship edges to trace the rule propagation path formed by the logical dependency relationship and the temporal triggering relationship, identify all rule nodes that are cascaded and trigger, and construct a complete rule chain.

[0013] Starting from the initial set of rule nodes, a graph traversal is performed along the rule relationship edges to expand the graph, tracing the rule propagation path formed by the logical dependencies and the temporal triggering relationships, identifying all cascaded triggering rule nodes, and constructing a complete rule chain, including: Extract the rule semantic features and trigger condition expressions of each rule node in the initial rule node set, and calculate the semantic neighborhood range of each rule node in the dynamic rule knowledge graph based on the rule semantic features; Within the semantic neighborhood, candidate successor rule nodes connected to the initial rule node set through rule relationship edges are identified, and the precondition constraints corresponding to logical dependencies and the time sequence constraints corresponding to time sequence triggering relationships are parsed from the rule relationship edges. The state of the trigger condition expression of each rule node in the initial rule node set is used as the precondition input. Logical deduction is performed on the precondition constraint to generate a logical dependency satisfaction evaluation result. At the same time, the time sequence constraint is time sequence verified based on the trigger timestamp of each rule node in the initial rule node set to generate a time sequence trigger satisfaction evaluation result. The logical dependency satisfaction evaluation result and the timing trigger satisfaction evaluation result are jointly determined, and candidate successor rule nodes that meet the joint determination conditions are selected as cascaded trigger nodes. The cascaded triggering node is used as a new initial rule node set, and the semantic neighborhood range calculation and the candidate successor rule node identification are performed iteratively until no new cascaded triggering node is added. Summarize all the cascaded trigger nodes identified during the iteration process and the rule relationship edges they connect. Construct a directed graph structure based on the propagation direction of the rule relationship edges to form a complete rule chain.

[0014] A second aspect of this invention provides an image content compliance review system based on AI reasoning, comprising: The scene semantic unit is used to acquire the image to be reviewed and its associated multi-source context data, perform multi-granularity region segmentation and object relationship modeling on the image to be reviewed, extract the spatial topological relationship and semantic dependency relationship between object entities in the image, generate a structured scene graph representation, and perform context sensitivity weighting on the nodes in the structured scene graph representation based on the multi-source context data to form a context-aware scene semantic representation. The compliance assessment unit is used to input the scene semantic representation into the reasoning module, and perform explicit violation detection and implicit risk prediction in parallel based on the spatial topological relationship and semantic dependency relationship in the scene semantic representation to generate a multi-dimensional compliance assessment result. The decision reasoning unit is used to establish a dynamic rule knowledge graph, perform reasoning path search in the dynamic rule knowledge graph based on the multi-dimensional compliance assessment results, identify the triggered rule chains and conflict relationships between rules, generate hierarchical review decisions, and extract the rule nodes in the reasoning path and the corresponding evidence nodes in the scene semantic representation to construct the causal explanation chain of the review decision. The dynamic update unit is used to execute the hierarchical review decision, collect manual review feedback and user appeal data, use the manual review feedback and user appeal data to incrementally update the rule entity relationships in the dynamic rule knowledge graph, and at the same time adjust the context sensitivity weight in the scene semantic representation according to the causal explanation chain.

[0015] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0017] This method constructs a structured scene graph through multi-granularity region segmentation and object relationship modeling, which can accurately capture the spatial topology and semantic dependencies between objects in the image. By combining multi-source contextual data to weight nodes, a context-aware scene semantic representation is formed, which effectively improves the accuracy of object association and semantic understanding in complex scenes and provides a comprehensive and fine-grained data foundation for subsequent compliance analysis.

[0018] By performing explicit violation detection and implicit risk prediction in parallel based on scene semantic representation, and making full use of spatial and semantic relationships for multi-dimensional evaluation, the system significantly improves the ability to identify direct violations and potential risk content, avoids missed or false judgments caused by a single detection mode, and enhances the coverage and accuracy of the review system.

[0019] By using a dynamic rule knowledge graph to search for reasoning paths and identify conflicting relationships, it can automatically match applicable rule chains and handle contradictions between rules. While generating hierarchical review decisions, it also constructs causal explanation chains, realizing transparency and interpretability of the review process. This provides a clear logical basis for decision-making and facilitates manual review and system optimization.

[0020] By leveraging feedback from manual reviews and user complaints, the system incrementally updates and retrospectively adjusts the rule knowledge graph and scenario representation weights, enabling it to continuously learn and evolve. This allows it to keep pace with changes in compliance requirements and the semantic environment, continuously optimizing the accuracy and efficiency of the review process, thus forming a closed-loop self-improvement mechanism. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the image content compliance review method based on AI reasoning, as described in an embodiment of the present invention. Figure 2 This is a flowchart of the cascaded trigger recognition process based on joint determination and iterative expansion in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0024] Figure 1 This is a flowchart illustrating the image content compliance review method based on AI reasoning, as described in an embodiment of the present invention. Figure 1 As shown, the image content compliance review methods based on AI reasoning include: The process involves acquiring images to be reviewed and their associated multi-source context data, performing multi-granularity region segmentation and object relationship modeling on the images to be reviewed, extracting spatial topological relationships and semantic dependencies between object entities in the images, generating a structured scene graph representation, and applying context-sensitive weighting to the nodes in the structured scene graph representation based on the multi-source context data to form a context-aware scene semantic representation. The scene semantic representation is input into the reasoning module, and explicit violation detection and implicit risk prediction are performed in parallel based on the spatial topological relationship and semantic dependency relationship in the scene semantic representation to generate multi-dimensional compliance assessment results. A dynamic rule knowledge graph is established. Based on the multi-dimensional compliance assessment results, reasoning path search is performed in the dynamic rule knowledge graph to identify the triggered rule chains and conflict relationships between rules, generate hierarchical review decisions, and extract the rule nodes in the reasoning path and the corresponding evidence nodes in the scene semantic representation to construct a causal explanation chain for the review decision. The hierarchical review decision is executed, and manual review feedback and user appeal data are collected. The manual review feedback and user appeal data are used to incrementally update the rule entity relationships in the dynamic rule knowledge graph. At the same time, the context sensitivity weight in the scene semantic representation is adjusted backtrackingly according to the causal explanation chain.

[0025] In one optional implementation, the image to be reviewed undergoes multi-granularity region segmentation and object relationship modeling to extract spatial topological relationships and semantic dependencies between object entities in the image, generating a structured scene graph representation, including: A multi-scale feature pyramid is constructed, and the image to be reviewed is segmented into hierarchical regions using the multi-scale feature pyramid. Candidate region sets are generated at different scale levels, and cross-level consistency checks are performed on the candidate regions at each scale level to select a stable region set with scale invariance. Object entity identification and attribute labeling are performed on each region in the stable region set to obtain an object entity set containing object category, location coordinates and semantic attributes. The relative orientation and spatial distance between objects are calculated based on the location coordinates of each object entity in the object entity set to determine the spatial topological relationship. The semantic attributes of each object entity in the object entity set are mapped to a pre-constructed semantic knowledge space. The semantic distance and co-occurrence probability between object entities in the semantic knowledge space are calculated. The semantic distance is then spatially constrained and corrected in conjunction with the spatial topological relationship to generate a relation weight matrix representing semantic dependencies. A structured scene graph representation is constructed using the object entities in the object entity set as nodes and the spatial topological relationship and the relationship weight matrix as the attributes of the edges.

[0026] To address multi-granularity region segmentation and object relationship modeling of images under review, a multi-scale feature pyramid is first constructed to obtain feature representations of the image at different resolutions. A convolutional neural network is used as the backbone for feature extraction, extracting feature maps at multiple deep layers of the network to form a multi-scale feature set containing both low-level detailed features and high-level semantic features. Specifically, for an input size of... The images awaiting review are used to generate a five-layer feature pyramid through successive convolution and pooling operations. The dimensions of the feature maps in each layer are as follows: ,in The value ranges from 1 to 5. To enhance the semantic consistency of features at different levels, a feature fusion mechanism is introduced between adjacent layers of the pyramid. High-level features are upsampled to the same spatial size as low-level features, and then concatenated with low-level features along the channel dimension. This ensures that each layer of features contains both the original features at that scale and supplementary information from other scales.

[0027] At each level of the multi-scale feature pyramid, a region candidate network generates a set of candidate regions for that scale. The region candidate network sets multiple anchor boxes with different aspect ratios at each spatial location in the feature map, predicting the probability of an object existing at that location and the corrected offset of the bounding box for each anchor box. The anchor boxes are sorted according to the predicted probabilities, and anchor boxes with probabilities higher than a preset threshold are selected as candidate regions. For the ... Candidate region set generated by layer feature map Each candidate region From bounding box coordinates With confidence score This indicates that, due to the differences in object size ranges captured at different scale levels, cross-level consistency checks are required for candidate regions at each level to eliminate duplicate detections.

[0028] Cross-level consistency verification is achieved by calculating the spatial overlap of candidate regions at different scale levels. For two candidate regions from different levels... and The intersection-union ratio (IUR) is calculated as a measure of spatial overlap. When the IUR exceeds 0.7, the two candidate regions are considered to correspond to the same object entity. In the set of candidate regions with cross-level overlap, the candidate region with the highest confidence score is retained, while other overlapping regions are suppressed. After cross-level consistency verification, the candidate regions retained from each level are merged to form a stable region set. The regions in this set exhibit cross-scale detection consistency, effectively reducing detection instability caused by scale variations.

[0029] Object entity recognition and attribute labeling are performed on each region in the stable region set. The feature map region corresponding to each candidate region is converted into a fixed-size feature vector through RoI pooling. This feature vector is input into the classification and regression branches. The classification branch outputs the probability distribution of object categories through a fully connected layer and a softmax activation function, and selects the category with the highest probability as the object category label for that region. The regression branch further fine-tunes the bounding box coordinates to more accurately fit the actual boundary of the object. In addition to object category and location information, semantic attribute features of the object are also extracted, including visual attributes such as color, texture, and pose, as well as text semantic embeddings corresponding to the object category obtained from a pre-trained language model. The object category, bounding box coordinates, and semantic attributes are integrated to form an object entity representation. ,in For category labels, For bounding box coordinates, These are attribute feature vectors. All object entities constitute the object entity set. ,in This represents the total number of objects detected.

[0030] Spatial topological relationships are determined based on the position coordinates of each object in the object entity set. For any two objects in the object entity set... and Extract the coordinates of the center point of its bounding box. and Calculate the relative azimuth angle between the two objects. Spatial distance The relative azimuth angle is determined by the angle between the line connecting the two center points and the horizontal axis, dividing the 360 ​​degrees into eight azimuth intervals, corresponding to up, down, left, right, and four diagonal directions. Spatial distance is measured using the Euclidean distance between the center points of the bounding boxes, normalized to the image diagonal length to eliminate the influence of image size. In addition to the center point distance, the minimum edge distance between the bounding boxes of two objects is calculated as a supplementary measure of spatial proximity. When the bounding boxes of two objects overlap, this distance is negative to indicate the degree of overlap. The relative azimuth, center distance, and edge distance are integrated to form a feature vector describing spatial topological relationships. .

[0031] To establish semantic dependencies between objects, the semantic attributes of object entities are mapped to a pre-constructed semantic knowledge space. This semantic knowledge space is pre-trained on a large-scale visual-language dataset, embedding object categories and attributes into a unified semantic vector space, ensuring that semantically related concepts are close together within this space. For objects... and Obtain its attribute feature vector and In the semantic knowledge space representation, the cosine similarity between the two is calculated as a semantic distance metric. Simultaneously, the co-occurrence probability of two object categories is queried from a pre-trained visual relationship knowledge base. This probability reflects the frequency with which two types of objects appear simultaneously in a real-world scenario. The product of semantic distance and co-occurrence probability serves as the initial semantic association strength.

[0032] Relying solely on semantic similarity can lead to unreasonable dependencies because semantically similar objects are spatially far apart and have no actual interaction. Therefore, spatial constraints are introduced to correct for semantic distance. Spatial constraint coefficients are defined. As the center distance The spatial constraint coefficient decreases as the spatial distance between two objects increases, thus reducing the semantic association strength. Specifically, the spatial constraint coefficient is calculated using a Gaussian kernel function, where the bandwidth parameter of the kernel function is adaptively adjusted according to the image size. Multiplying the semantic association strength by the spatial constraint coefficient yields the semantic dependency strength corrected for spatial constraints. For all object pairs, calculate the semantic dependency strength, thus forming... dimensional relational weight matrix The first in the matrix Line number The elements of the column are .

[0033] When generating a structured scene graph representation, each object entity in the object entity set is treated as a node in the graph, carrying the object's category, location, and attribute information. For object pairs whose weight values ​​in the relation weight matrix exceed a preset threshold, directed edges are established between the corresponding nodes, and the edge attributes include spatial topological relation feature vectors. With semantic dependency strength Spatial topological relationships describe the geometric layout between objects, while semantic dependencies characterize the semantic interaction patterns between objects. Together, they constitute a complete representation of edges in the scene graph. For object pairs with overlapping bounding boxes, occlusion markers are added to the edge attributes to indicate which object is in the foreground and which is in the background. The resulting structured scene graph contains a set of nodes, a set of edges, and a set of attributes for nodes and edges, fully representing the spatial structure and semantic organization of the image to be reviewed, providing a structured input representation for subsequent compliance reasoning.

[0034] In one optional implementation, a multi-scale feature pyramid is constructed. The image to be reviewed is then segmented into hierarchical regions using the multi-scale feature pyramid. Candidate region sets are generated at different scale levels, and cross-level consistency checks are performed on the candidate regions at each scale level. A set of stable regions with scale invariance is selected, including: The images to be reviewed are downsampled step by step to generate multiple image layers with decreasing resolution. Semantic features and structural features are extracted from each image layer. A bidirectional feature transfer path is established between adjacent image layers. The semantic features and structural features of different image layers are fused across layers through the bidirectional feature transfer path to construct a multi-scale feature pyramid. Region extraction operations are performed at each image level of the multi-scale feature pyramid to generate a candidate region set corresponding to each image level. Cross-level spatial alignment is performed on the candidate region set, projecting candidate regions from different image levels onto a unified reference coordinate system. Cross-level candidate region pairs with overlapping spatial positions in the unified reference coordinate system are identified. The feature response consistency of each cross-level candidate region pair is calculated. When the feature response consistency meets a preset stability condition, the corresponding cross-level candidate region pairs are merged into a single stable region. The single stable regions are then summarized to form a stable region set.

[0035] In practical applications, when performing stepwise downsampling on images to be reviewed, the original input image is used as the first-level image, maintaining its original resolution. Subsequent downsampling operations reduce the width and height of the current level image to half their original size each time, thus constructing a sequence of image levels with multiple different resolutions. In a specific implementation scenario, if the original image resolution is 1024 pixels by 1024 pixels, after one downsampling, a second-level image of 512 pixels by 512 pixels is obtained; after another downsampling, a third-level image of 256 pixels by 256 pixels is obtained, and so on, until the minimum side length of the image is reduced to 32 pixels or 64 pixels, ensuring that sufficient semantic information is retained at the smallest level for subsequent analysis.

[0036] After obtaining each image layer, semantic and structural features are extracted for each layer. Semantic feature extraction utilizes a pre-trained deep convolutional neural network to perform forward propagation on the images at each layer, obtaining feature activation responses from deeper layers of the network. These responses can capture semantic category information, attribute features, and high-level abstract concepts of objects in the image. Structural feature extraction focuses on local geometric information such as edges, textures, and corners of the image. This can be obtained through traditional image processing operators such as the Canny edge detection operator, Gabor filter banks, or convolutional features from shallow layers of the network. These structural features play an important role in recognizing the shape contours, spatial layout, and detailed textures of objects.

[0037] To enable information exchange between different image layers, a bidirectional feature transfer path is established between adjacent image layers. The bottom-up feature transfer path transfers high-resolution structural features from lower-level images to higher-level images, providing more refined spatial localization information and edge details. The top-down feature transfer path transfers semantic features from higher-level images to lower-level images, injecting richer semantic understanding into the feature representations of lower levels. When implementing bidirectional transfer, due to the difference in image resolution between adjacent layers, the spatial dimension of the transferred features needs to be adjusted. For bottom-up transfer, pooling operations or strided convolutions can be used to downsample lower-level features to match their spatial dimensions with the receiving layer. For top-down transfer, upsampling operations such as bilinear interpolation or transposed convolutions are used to restore the spatial resolution of higher-level features to be consistent with the receiving layer.

[0038] When implementing cross-level feature fusion via a bidirectional feature transfer path, the transferred features from adjacent levels are fused with the original features extracted at each image level. The fusion method can employ element-wise addition, element-wise multiplication, or transformation of the concatenated features using learnable convolutional layers. The fused features simultaneously contain the original information of that level, fine structural information from lower levels, and abstract semantic information from higher levels. This integration of multi-source information makes the feature representation of each image level more comprehensive, thus constructing a complete multi-scale feature pyramid.

[0039] After constructing the multi-scale feature pyramid, region extraction is performed separately for each image level within the pyramid. Region extraction can utilize selective search algorithms, region proposal networks, or dense sliding windows to generate candidate regions. In a specific implementation, a region proposal network is used to perform foreground / background classification and bounding box regression on the fused feature map of each image level, generating a candidate region set corresponding to that level. Due to the resolution differences between image levels, candidate regions extracted at low-resolution levels tend to cover larger-scale objects in the image, while candidate regions extracted at high-resolution levels can accurately locate smaller-scale objects or local details of large objects. By performing region extraction independently at each level, comprehensive coverage of objects at different scales can be achieved.

[0040] To select scale-invariant stable regions from the large number of candidate regions generated at each level, cross-level spatial alignment is required. Since the coordinates of candidate regions at each level are defined in their respective image coordinate systems, the coordinates of all candidate regions at all levels are first projected onto the coordinate system of the original image as a unified reference coordinate system. Specifically, during the projection process, for the... The coordinates at each image level are The candidate regions, whose corresponding coordinates in the unified reference coordinate system are calculated as follows: ,in For the first The scale factor of each level relative to the original image.

[0041] Under a unified reference coordinate system, candidate region pairs from different levels that overlap spatially are identified. For any two candidate regions, the degree of spatial overlap is evaluated by calculating their intersection-union ratio (IU). The IU is defined as the ratio of the intersection area to the union area of ​​two candidate regions. When the IU exceeds a set threshold, such as 0.5 or 0.6, the two candidate regions are determined to constitute a cross-level candidate region pair, indicating that they spatially point to the same image region but come from different scale levels.

[0042] For each identified cross-level candidate region pair, their feature response consistency is further calculated. Feature response consistency measures the semantic similarity of the features extracted from the two candidate regions. Specifically, for each candidate region in the cross-level candidate region pair, the corresponding region feature vectors from the fused feature pyramid are extracted. Region of interest pooling can be used to convert the feature map regions corresponding to the candidate regions into fixed-dimensional feature vectors. Then, the cosine similarity between the two feature vectors is calculated; this cosine similarity value is the quantitative indicator of feature response consistency.

[0043] A preset stability condition is used to determine whether cross-level candidate region pairs correspond to a stable representation of the same real object. This condition includes constraints in two dimensions: the spatial dimension requires the intersection-union ratio to exceed a preset spatial overlap threshold, and the feature dimension requires the cosine similarity of feature response consistency to exceed a preset feature similarity threshold, such as 0.8. Only cross-level candidate region pairs that simultaneously satisfy both of these constraints are considered to have stable representations at different scales, representing a consistent manifestation of the same object at different observation scales.

[0044] When a pair of cross-level candidate regions satisfies the stability condition, the pair is merged into a single stable region. Several strategies can be employed for merging. One strategy is to take a weighted average of the coordinates of the two candidate regions in a unified reference coordinate system, with the weights determined based on the feature response intensity or detection confidence of their respective levels. Another strategy is to select the region with higher confidence among the two candidate regions as the representative region after merging. The merging process also incorporates feature information from different levels, enabling the single stable region to integrate multi-scale feature representations.

[0045] After verifying and merging the consistency of all cross-level candidate region pairs, all identified single stable regions are aggregated. For candidate regions that fail to find a corresponding relationship in cross-level matching, they are filtered based on their own detection confidence and feature response strength. Single-level candidate regions with higher confidence are also included in the stable region set to ensure comprehensive coverage of objects at various scales in the image. The final stable region set includes both high-confidence regions validated at multiple scales and significant object regions at a single scale, providing a reliable spatial localization foundation for subsequent object relationship modeling and scene graph construction.

[0046] In one optional implementation, the nodes in the structured scene graph representation are weighted by context sensitivity based on the multi-source context data to form a context-aware scene semantic representation, including: Time series features, geographic location features, and user interaction features are extracted from multi-source context data, and cross-modal semantic alignment is performed on the time series features, geographic location features, and user interaction features to generate a fused context semantic embedding vector. The semantic representation of each node in the structured scene graph representation is matched with the context semantic embedding vector for similarity, and the semantic activation strength of each node under the current context condition is calculated. Based on the semantic activation intensity, a node importance ranking sequence is constructed, and a node sensitivity weight is calculated according to the relative position of each node in the node importance ranking sequence. The node sensitivity weight is then applied to the feature representation of the corresponding node in the structured scene graph representation to generate a context-weighted node feature representation. The node feature representations and their associated edge relationships are aggregated by graph convolution to form a context-aware scene semantic representation.

[0047] In the process of applying context-sensitive weighting to nodes in a structured scene graph representation, it is necessary to fully utilize multi-source contextual data to enhance the accuracy of scene understanding. Images awaiting review are often not isolated entities; their compliance assessment requires considering multiple dimensions of information, including the image upload time, geographical location, and user historical behavior. First, features are extracted from the multi-source contextual data. Time-series features include the specific time of image upload, the time period in which it occurred, and the time distance from specific events. A time-coding network maps these discrete time information into a continuous vector representation. Geographical location features encompass the image's shooting location, the administrative division information of the upload location, and the location's sensitivity level. A geocoding network converts heterogeneous geographic information such as latitude and longitude coordinates and location names into a unified spatial embedding representation. User interaction features involve the uploading user's historical behavior patterns, account reputation score, recent interaction frequency, and content type preferences. A user profiling module abstracts this behavioral data into user feature vectors.

[0048] Since time-series features, geographic location features, and user interaction features originate from different data sources and have heterogeneous representations, cross-modal semantic alignment is required to bridge the semantic gap between different feature spaces. A cross-modal alignment network is constructed, comprising encoding branches for each feature type and a shared semantic projection layer. Extracted time-series features are input into the time encoding branch, where a multilayer perceptron maps the time features to an intermediate semantic space. Geographic location features are input into the spatial encoding branch, where a graph neural network captures the relationships between geographic locations. User interaction features are input into the behavior encoding branch, where an attention mechanism is used to weight and aggregate different behavior dimensions. The output vectors of the three encoding branches are fed into the shared semantic projection layer, which uses contrastive learning to shorten the distance between relevant contextual features and widen the distance between irrelevant features, thus achieving cross-modal semantic alignment. The aligned time feature vector, geographic feature vector, and user feature vector are concatenated and then subjected to a nonlinear transformation through a fusion layer to generate a fused contextual semantic embedding vector. This vector comprehensively represents the complete context of the current image.

[0049] After obtaining the contextual semantic embedding vector, it needs to be associated with the semantics of each node in the structured scene graph representation. Each node in the structured scene graph represents an object entity in the image, and each node has its own semantic representation. This representation is typically composed of the object's category characteristics, visual characteristics, attribute characteristics, etc.

[0050] semantic representation of computing nodes With context semantic embedding vector The similarity between them is measured using cosine similarity, and the similarity is calculated as follows: This similarity value reflects the degree of association between a specific object node and its current context. For example, an image containing alcoholic beverages uploaded at night will have a higher similarity between the alcoholic beverage node and the nighttime time context, because this combination carries a higher risk in that specific scenario. To further capture the activation effect of context on node semantics, a non-linear activation function is introduced to transform the similarity, and the semantic activation strength is calculated as follows: ,in The weight matrix is ​​a learnable matrix. For bias terms, For activation function, symbol This represents a vector concatenation operation. The activation strength considers not only similarity information but also integrates node-specific features and contextual features, enabling a more accurate characterization of a node's salience within the current context.

[0051] Based on the calculated semantic activation strength of each node, a node importance ranking sequence is constructed, and all nodes in the scene graph are ranked according to their semantic activation strength. The nodes are sorted from highest to lowest to form an ordered sequence. Nodes with higher rankings are considered more important or riskier in the current context and require greater attention during subsequent reviews. Node sensitivity weights are calculated based on their relative positions in the sorted sequence, and a position decay function is used to ensure that important nodes receive higher weights.

[0052] The node sensitivity weight is calculated as follows: ,in This represents the index of the node's position in the sorted sequence. This represents the total number of nodes in the scene graph. The temperature parameter is used to control the steepness of the weight distribution. This calculation method gives significantly higher weights to top-ranked nodes while ensuring that the sum of all weights is 1, satisfying the properties of a probability distribution. The calculated node sensitivity weights are then used... The original feature representation applied to the corresponding node is weighted using element-wise multiplication to obtain a context-weighted node feature representation. This weighting operation modulates node features based on contextual information, strengthening the features of nodes that are more important in the current context, while relatively suppressing the features of irrelevant nodes.

[0053] After obtaining the context-weighted node feature representations, it is also necessary to consider the impact of the relationships between nodes on the scene semantics. Edge relationships in the structured scene graph describe the spatial, interactive, or semantic relationships between objects, and the information carried by these edge relationships is crucial for understanding the overall scene semantics. Graph convolution aggregation operations are performed on the context-weighted node feature representations and their associated edge relationships, employing a graph convolutional network architecture to achieve the propagation and fusion of information on the graph structure.

[0054] For nodes The graph convolution update process is as follows: ,in Indicates the first Layer node characteristics, Represents a node The set of neighboring nodes, and They are nodes and The degree, Indicates the connection node and edge features, and This is a learnable weight matrix.

[0055] This convolutional process allows each node to retain not only its own context-weighted features but also information from neighboring nodes, thereby capturing the contribution of relationships between objects to scene understanding. By stacking multiple layers of graph convolutional operations, information can propagate across a larger graph structure, effectively capturing dependencies between distant nodes. The output of the final graph convolutional layer is a context-aware scene semantic representation. This representation includes both the context-weighted features of each object node and the structural information of relationships between objects, comprehensively reflecting the scene semantics of the image under specific contextual conditions.

[0056] In practical applications, the quality and completeness of contextual data directly affect the accuracy of the weighting effect. When data for certain contextual dimensions is missing, default values ​​are used to fill the gaps, or estimates based on historical statistics are used to complete the data. Temperature parameter The selection needs to be optimized based on the specific application scenario; larger ones A smaller value will cause the weight distribution to be more concentrated in a few high-ranking nodes, which is suitable for scenarios where key objects need to be focused on; The value makes the weight distribution more even, suitable for scenarios that need to comprehensively consider all objects. The number of layers in the graph convolution also needs to be selected according to the average diameter of the scene graph. Usually, 2 to 3 layers can cover most of the paths between nodes. Too deep a network will lead to oversmoothing. Through this series of operations, the final generated context-aware scene semantic representation can effectively support subsequent compliance reasoning and review decisions.

[0057] In one optional implementation, the scene semantic representation is input into the inference module, and explicit violation detection and implicit risk prediction are performed in parallel based on the spatial topological relationships and semantic dependencies in the scene semantic representation, generating multi-dimensional compliance assessment results including: Extract object location distribution features corresponding to spatial topological relationships and object interaction pattern features corresponding to semantic dependencies from scene semantic representation, and input the object location distribution features into the explicit violation detection branch and the object interaction pattern features into the implicit risk prediction branch. In the explicit violation detection branch, the location distribution features of the object are matched with rules based on a pre-built violation rule base to identify explicit violation objects that violate explicit prohibitions and generate explicit violation detection results; In the implicit risk prediction branch, a behavioral association graph between objects is constructed based on the object interaction pattern characteristics. By analyzing the abnormal association patterns in the behavioral association graph, objects with potential hidden risks that could lead to violations are identified, and implicit risk prediction results are generated. The explicit violation detection results are integrated with the implicit risk prediction results, and the explicit violation objects and the implicit risk objects are respectively assigned violation severity scores and risk tendency scores to generate a multi-dimensional compliance assessment result that includes the violation severity scores and the risk tendency scores.

[0058] After the scene semantic representation is input into the inference module, it needs to separate and extract the different types of features it contains. The object location distribution features corresponding to spatial topological relationships mainly reflect the spatial layout patterns of various object entities in the image, including the absolute position coordinates, relative positional relationships, spatial density, and geometric distances between object bounding boxes. These features are obtained by aggregating the spatial attributes of nodes in the scene graph representation. Specifically, for each object node in the scene graph, the coordinates of its center point in the original image coordinate system are extracted. Boundary box dimensions And calculate the Euclidean distance between the object and other object nodes. This process forms a distance matrix between objects. Simultaneously, the distribution density of object nodes within the image region is statistically analyzed. The image is divided into a uniform grid, and the number of objects within each grid is calculated, creating a spatial density heatmap. These locational distribution features are organized into feature vectors, which serve as input to the explicit violation detection branch.

[0059] Semantic dependency relationships, corresponding to object interaction patterns, characterize the semantic associations and behavioral connections between object entities, including action relationships, attribute modification relationships, and logical subordination relationships. During extraction, relationship type labels connecting object nodes are obtained from the edge information of the scene graph representation, such as semantic relationships in the form of triples like "person-holding-object" and "background-containing-object". For each relationship type, its frequency and distribution pattern in the scene graph are statistically analyzed to construct a semantic association matrix between objects. For relationships involving actions, attribute information such as the directionality and intensity of the actions are extracted; for modification-type relationships, the sentiment tendency and sensitivity level of the modifiers are extracted. These interaction pattern features are encoded into high-dimensional vectors and input into the implicit risk prediction branch for processing.

[0060] After receiving the object location distribution features, the explicit violation detection branch initiates a rule-based matching process. The pre-built violation rule base stores explicit prohibitions, organized in a structured rule format. Each rule includes elements such as the violating object category, spatial layout constraints, and contextual limitations. For example, for a specific sensitive object, the rule states that "when object A appears in the center area of ​​an image and its proportion exceeds a certain threshold, it is considered a violation." During rule matching, each rule in the rule base is traversed, comparing the spatial layout constraints defined in the rule with the input object location distribution features. During comparison, it checks whether any object satisfies the rule constraints, specifically determining whether the object's position coordinates fall within the spatial range specified by the rule, whether the object's size exceeds the threshold set by the rule, and whether the object's relative distance to other specific objects violates the rule requirements. When an object fully satisfies all the constraints of a rule, it is marked as an explicit violation object, and the triggered rule number and violation type are recorded. After traversal, all marked explicit violation objects are summarized to form the explicit violation detection result, which includes detailed information such as the identifier of each violation object, the triggering rule, and the violation location coordinates.

[0061] After receiving the object interaction pattern features, the implicit risk prediction branch initiates the behavior association graph construction process. The behavior association graph uses object entities as nodes and semantic dependencies between objects as edges, forming a directed graph structure. During construction, all object interaction pattern features extracted from the scene semantic representation are mapped into the graph. For each pair of objects with a semantic relationship, a directed edge is established between the corresponding nodes, and the edge attributes record information such as relationship type and relationship strength. Relationship strength is calculated based on the salience of the semantic dependency in the scene graph; relationships with higher salience are assigned greater weight. After the graph construction is complete, the abnormal association pattern identification process is initiated. Abnormal association patterns refer to rare or unreasonable combinations of objects and interaction methods in normal scenarios. Although these patterns do not directly violate explicit rules, they have the potential risk of causing violations.

[0062] During the identification process, a graph pattern mining algorithm is used to search for predefined abnormal pattern templates in the behavioral association graph. The template library stores high-risk interaction patterns summarized from historical data, such as "specific object combinations co-occurring with specific spatial relationships" and "abnormally frequent association edges between objects." Simultaneously, anomaly detection algorithms are used to identify graph structures deviating from normal distributions. By calculating the statistical deviations of graph features such as node degree distribution, subgraph density, and path length, abnormally clustered or abnormally sparse local structures are discovered. When an abnormal pattern is matched or a structural anomaly is detected, the involved object nodes are extracted and marked as latent risk objects, recording the triggered abnormal pattern type and degree of anomaly. After identification is complete, all latent risk objects are summarized to form a latent risk prediction result, which includes risk object identifiers, associated abnormal patterns, and involved interaction relationship chains.

[0063] After the explicit violation detection results and implicit risk prediction results are generated, a dual-result fusion process is initiated. During the fusion process, a unified evaluation framework needs to be established for the two different types of detected objects. For explicit violations, a violation severity score is assigned based on the severity level of the triggering rule. Each rule in the violation rule library is pre-configured with a severity level parameter, and the level classification is determined based on factors such as the social impact and legal consequences of the violation. When calculating the severity score, the level parameter of the triggering rule for the object is extracted, while also considering the salience of the violating object in the image. Salience is quantified by factors such as object size proportion and positional importance. The final violation severity score is determined by combining the rule level and object salience. For implicit risk objects, a risk tendency score is assigned based on the risk tendency of the identified abnormal patterns. Risk tendency reflects the probability that a combination of objects will cause an actual violation. When calculating the score, the statistical probability of an abnormal pattern transforming into an actual violation in historical data is extracted, combined with the matching strength of the pattern in the current scene. The matching strength is measured by the similarity between the abnormal pattern features and the actual image features. At the same time, considering the connectivity of objects in the behavior association graph, the higher the connectivity of an object, the greater its impact on the overall scenario, and the higher its risk tendency score.

[0064] After scoring is completed, a multi-dimensional compliance assessment result is constructed. This result is organized in a structured data format, containing both object-level assessment information and image-level comprehensive assessment information. The object-level assessment information lists all marked explicit violations and implicit risk objects. Each object records detailed attributes such as its category, location, corresponding violation severity score or risk tendency score, triggered rules, or anomaly patterns. The image-level comprehensive assessment information is generated by aggregating the scores of all objects, calculating the overall violation risk index for the image. The violation risk index comprehensively considers factors such as the maximum severity score of explicit violations, the cumulative score of all violations, and the risk tendency score distribution of implicit risk objects, calculated through weighted summation or a non-linear mapping function. The assessment result also includes the location information of each object in the original image, facilitating precise location of problem areas during subsequent review decisions. Simultaneously, the assessment result records the execution time and resource consumption of the two detection branches, providing reference data for system performance optimization. Once generated, the multi-dimensional compliance assessment result serves as input for subsequent inference path search and hierarchical review decisions, supporting decision-making throughout the entire review process.

[0065] In one optional implementation, a dynamic rule knowledge graph is established. Based on the multi-dimensional compliance assessment results, reasoning path searches are performed in the dynamic rule knowledge graph to identify triggered rule chains and conflict relationships between rules, generating tiered review decisions, including: A dynamic rule knowledge graph containing rule nodes and rule relationship edges is constructed, wherein the rule relationship edges represent the logical dependencies and temporal triggering relationships between rules, and the dynamic rule knowledge graph is updated in real time to reflect the dynamic changes of the rule base. Using the severity score and risk propensity score from the multi-dimensional compliance assessment results as query conditions, the initial set of rule nodes that match the severity score and risk propensity score are located in the dynamic rule knowledge graph. Starting from the initial set of rule nodes, a graph traversal expansion is performed along the rule relationship edges to trace the rule propagation path formed by the logical dependency relationship and the temporal triggering relationship, identify all rule nodes that are cascaded and trigger, and construct a complete rule chain.

[0066] like Figure 2 As shown, the method includes: The dynamic rule knowledge graph is constructed using a graph database storage structure. Rule nodes have the following attribute fields: unique identifier, rule type label, rule name, rule description text, rule priority value, rule status identifier, creation timestamp, and last update timestamp. The rule type labels for rule nodes include three categories: content review rules, behavior monitoring rules, and risk warning rules. The rule status identifier includes three values: enabled, disabled, and under testing. Rule relationship edges have the following attribute fields: edge unique identifier, edge type label, source node identifier, target node identifier, edge weight value, constraint expression, and creation timestamp. Edge type labels include two types: logical dependency relationships and temporal triggering relationships. Logical dependency relationships indicate that the triggering of the target rule node depends on the judgment result of the source rule node, while temporal triggering relationships indicate that the triggering of the target rule node must occur within a specific time window after the source rule node is triggered. Constraint expressions are stored in key-value pairs, where the key is the constraint type (including preconditions, time intervals, and trigger counts), and the value is the corresponding constraint parameter. Real-time updates to the dynamic rule knowledge graph are achieved through a rule change monitoring mechanism. When the rule base receives rule addition, rule modification, or rule deletion operations, the monitoring module captures the change events and parses the change content. The change content is then converted into node insertion, attribute update, and node deletion operation instructions for the graph database. The consistency between the graph state and the rule base state is ensured through transaction commit.

[0067] When using the severity score and risk propensity score from the multi-dimensional compliance assessment results as query conditions, the severity score and risk propensity score are first mapped to different ranges. The severity score is segmented into three ranges: 0-30 (low risk), 30-60 (medium risk), and 60-100 (high risk). The risk propensity score is segmented into three ranges: 0-40 (conservative), 40-70 (neutral), and 70-100 (aggressive). When locating the initial set of rule nodes in the dynamic rule knowledge graph, a graph query statement is constructed. The query conditions include: the rule type label of the rule node matching the assessment object type; the rule priority value of the rule node falling within the priority range corresponding to the severity score range; and the rule status of the rule node being enabled. The correspondence between the rule priority value and the severity score range is as follows: low risk range corresponds to priority 1-3, medium risk range to priority 4-6, and high risk range to priority 7-10. The query results return a set of rule nodes that satisfy all query conditions as the initial set of rule nodes. Each rule node carries a node identifier, rule name, rule priority value, and rule description text for subsequent processing.

[0068] When expanding the graph by traversing from the initial set of rule nodes, the semantic features and triggering condition expressions of each rule node in the initial set are extracted. The semantic features are obtained by text vectorization encoding of the rule description text. The encoding uses a pre-trained language model's encoding layer to output a fixed-dimensional feature vector, with the feature vector dimension set to 768 dimensions. Based on the semantic features, the semantic neighborhood range of each rule node in the dynamic rule knowledge graph is calculated. The semantic neighborhood range is determined by calculating the similarity between the feature vector of the rule node and the feature vectors of all other rule nodes in the graph. The similarity calculation uses vector dot product normalization, and the similarity value ranges from -1 to 1. When the similarity is greater than 0.6, the corresponding rule node is included in the semantic neighborhood range. The triggering condition expressions are extracted from the rule node attributes and stored in the form of logical judgment statements, including judgment conditions, judgment thresholds, and logical operators.

[0069] When identifying candidate successor rule nodes within the semantic neighborhood, the graph database is queried to find all adjacent rule nodes connected to the initial set of rule nodes via rule relationship edges. The filtering criteria are that the node identifiers of adjacent rule nodes are within the semantic neighborhood and the edge type label of the rule relationship edge is either logical dependency or temporal triggering relationship. From the filtered rule relationship edges, the precondition constraints corresponding to logical dependencies and the temporal constraints corresponding to temporal triggering relationships are parsed. The precondition constraint is extracted from the constraint expression of the edge attribute field, with the key being the value of the precondition. This value is stored in Boolean expression form, containing the judgment result state of the source rule node that must be satisfied. The temporal constraint is extracted from the constraint expression, with the key being the value of the time interval. This value is stored in time range form, including the minimum and maximum time intervals, with the time unit being seconds. The default value for the minimum time interval is 0, and the default value for the maximum time interval is 3600.

[0070] When the satisfaction status of the trigger condition expressions of each rule node in the initial rule node set is used as the precondition input, the satisfaction status of the trigger condition expression is obtained by comparing the judgment condition in the trigger condition expression with the actual attribute value of the current audit object. The comparison result is either satisfied or not satisfied. When performing logical deduction on the precondition constraints, the Boolean expression in the precondition constraint is logically matched with the satisfaction status of the trigger condition expression. The logical deduction result is true when the Boolean expression requires the judgment result of the source rule node to be true and the satisfaction status of the trigger condition expression is satisfied; the logical deduction result is true when the Boolean expression requires the judgment result of the source rule node to be false and the satisfaction status of the trigger condition expression is not satisfied; otherwise, the logical deduction result is false.

[0071] When generating the logical dependency satisfaction evaluation result, candidate successor rule nodes whose logical deduction results are true are marked as logically dependently satisfied. The result includes the candidate successor rule node identifier, the logical dependency satisfaction flag, and the deduction path description. When performing time-series validation on time-series constraints based on the trigger timestamps of each rule node in the initial rule node set, the trigger timestamp is obtained from the most recently updated timestamp field in the rule node attributes. Time-series validation is achieved by calculating the time difference between the current system time and the trigger timestamp and comparing it with the time range of the time-series constraints. The time-series validation passes when the time difference is between the minimum and maximum time intervals. When generating the time-series trigger satisfaction evaluation result, candidate successor rule nodes that pass the time-series validation are marked as time-series trigger satisfied. The result includes the candidate successor rule node identifier, the time-series trigger satisfaction flag, and the time difference value.

[0072] When jointly determining the logical dependency satisfaction evaluation result and the temporal trigger satisfaction evaluation result, the joint determination condition is that the candidate successor rule node simultaneously possesses both the logical dependency satisfaction identifier and the temporal trigger satisfaction identifier. When filtering out candidate successor rule nodes that meet the joint determination condition, the logical dependency satisfaction evaluation result and the temporal trigger satisfaction evaluation result are traversed. Nodes with the same candidate successor rule node identifier and both evaluation results are satisfied are marked as cascaded trigger nodes. Cascaded trigger nodes carry a node identifier, a trigger path identifier, and a trigger level value as extended attributes. The trigger path identifier records the complete path from the initial rule node set to the current node. The trigger level value represents the traversal depth of the current node from the initial rule node set. The trigger level value of the initial rule node set is 0, and it increments by 1 in each iteration.

[0073] When a cascaded trigger node is used as the new initial rule node set, the original initial rule node set is cleared, and the node identifier, rule name, rule priority value, rule description text, trigger path identifier, and trigger level value of the cascaded trigger node are assigned to the new initial rule node set. During iterative execution of semantic neighborhood range calculation and candidate successor rule node identification, the complete process of rule semantic feature extraction, semantic neighborhood range calculation, candidate successor rule node query, precondition constraint parsing, time sequence constraint parsing, logical deduction, time sequence verification, joint determination, and cascaded trigger node marking is repeatedly executed. The iteration terminates when the cascaded trigger node set is empty in the current iteration or the trigger level value reaches the preset maximum traversal depth. The default maximum traversal depth is 5 levels, which can be adjusted to 3 to 10 levels depending on the complexity of the rule chain.

[0074] When summarizing all cascaded trigger nodes identified during the iteration process, node identifiers, trigger path identifiers, and trigger level values ​​are extracted from the cascaded trigger node records of each iteration round, and duplicates are merged into a global set of cascaded trigger nodes. When extracting the rule-related edges connecting the cascaded trigger nodes, the graph database is queried based on the node identifier sequence recorded in the trigger path identifier to obtain the rule-related edges between adjacent nodes in the sequence. Edge attributes include a unique edge identifier, edge type label, source node identifier, target node identifier, and edge weight value. When constructing a directed graph structure based on the propagation direction of the rule-related edges, the cascaded trigger nodes are used as the vertex set, and the rule-related edges are used as the directed edge set. The direction of the edges points from the source node identifier to the target node identifier. An adjacency list data structure is constructed to store the directed graph, with the node identifier as the key and the list of all outgoing target node identifiers of that node as the value. When a complete rule chain is formed, the directed graph is topologically sorted. The sorting result is a sequence of node identifiers arranged in ascending order of trigger level value. Each node in the sequence carries a node identifier, rule name, rule priority value, trigger level value, a list of predecessor node identifiers, and a list of successor node identifiers. The rule chain is output in this sequence form for subsequent conflict detection and review decision generation.

[0075] In one optional implementation, starting from the initial set of rule nodes, a graph traversal expansion is performed along the rule relationship edges to trace the rule propagation path formed by the logical dependencies and the temporal triggering relationships, identifying all cascaded triggering rule nodes and constructing a complete rule chain, including: Extract the rule semantic features and trigger condition expressions of each rule node in the initial rule node set, and calculate the semantic neighborhood range of each rule node in the dynamic rule knowledge graph based on the rule semantic features; Within the semantic neighborhood, candidate successor rule nodes connected to the initial rule node set through rule relationship edges are identified, and the precondition constraints corresponding to logical dependencies and the time sequence constraints corresponding to time sequence triggering relationships are parsed from the rule relationship edges. The state of the trigger condition expression of each rule node in the initial rule node set is used as the precondition input. Logical deduction is performed on the precondition constraint to generate a logical dependency satisfaction evaluation result. At the same time, the time sequence constraint is time sequence verified based on the trigger timestamp of each rule node in the initial rule node set to generate a time sequence trigger satisfaction evaluation result. The logical dependency satisfaction evaluation result and the timing trigger satisfaction evaluation result are jointly determined, and candidate successor rule nodes that meet the joint determination conditions are selected as cascaded trigger nodes. The cascaded triggering node is used as a new initial rule node set, and the semantic neighborhood range calculation and the candidate successor rule node identification are performed iteratively until no new cascaded triggering node is added. Summarize all the cascaded trigger nodes identified during the iteration process and the rule relationship edges they connect. Construct a directed graph structure based on the propagation direction of the rule relationship edges to form a complete rule chain.

[0076] After obtaining the initial set of rule nodes, semantic features of each rule node are extracted. Each rule node carries attribute information such as rule type identifier, constraint description, and trigger threshold parameter in the dynamic rule knowledge graph. By performing structured parsing on these attribute information, the core entity type, constraint relationship type, threshold value, and other elements involved in the rule are extracted as semantic features. For the extraction of trigger condition expressions, the logical expressions stored in the rule nodes are converted into a computable form. This expression typically contains multiple sub-conditions connected by logical operators, such as "object category is contraband and confidence level is greater than 0.85", where the sub-conditions "object category is contraband" and "confidence level is greater than 0.85" are connected by a logical AND operator. Each sub-condition is further decomposed into three parts: judgment variable, comparison operator, and threshold parameter, forming a standardized expression structure.

[0077] When calculating the semantic neighborhood of a rule node based on the extracted semantic features, a rule semantic vector space is first constructed. Discrete features such as entity types and constraint types involved in the rule are vectorized and encoded using one-hot encoding or embedded vector representation. Continuous features such as thresholds are directly preserved in numerical form. The discrete feature vectors are then concatenated with the continuous features to form a comprehensive semantic vector. For a given rule node in the initial set of rule nodes, the similarity between its semantic vector and the semantic vectors of all other rule nodes in the dynamic rule knowledge graph is calculated using cosine similarity. The set of rule nodes with similarity scores higher than a preset threshold is selected as the semantic neighborhood of that initial rule node. This preset threshold can be dynamically adjusted according to the size of the rule base. When the rule base contains a large number of rules, the threshold is appropriately increased to narrow the neighborhood range and avoid excessive computational load in subsequent traversals.

[0078] Within a defined semantic neighborhood, candidate successor rule nodes directly connected to the initial set of rule nodes are identified by querying rule relationship edges in the dynamic rule knowledge graph. Rule relationship edges carry relationship type labels, including "logical dependency" and "temporally triggered." For relationship edges labeled "logical dependency," precondition constraints are parsed from the edge attributes. This constraint typically represents "the successor rule can be triggered when the preceding rule meets a specific state," such as "rule B can only be activated when rule A is determined to be in violation," where "rule A is determined to be in violation" constitutes the precondition constraint. For relationship edges labeled "temporally triggered," temporal constraints are extracted from the edge attributes. This constraint defines the triggering time relationship between the predecessor and successor rules, such as "a delay time window after rule A is triggered." "Internal rule B can be triggered", where the time window is... It is stored as a temporal constraint parameter in the relation edge.

[0079] The initial set of rule nodes uses the satisfaction status of their trigger condition expressions as precondition inputs for logical dependency deduction. Each rule node's trigger condition expression, after matching with the scene semantic representation, yields a Boolean value judgment result, indicating whether the current rule's trigger condition is satisfied by the scene content. For a candidate successor rule node, all predecessor rule nodes connected to it via logical dependency edges are obtained, and the trigger states of these predecessor rule nodes are extracted. Logical deduction expressions are constructed based on the precondition constraints defined in the logical dependency edges; for example, if a successor rule's precondition constraint is "rule..." Triggered and rules If "not triggered", the triggering state of the corresponding predecessor rule is substituted into the logical expression for calculation. If the expression is true, the logical dependency satisfaction evaluation result is "satisfied"; otherwise, it is "not satisfied". For cases where multiple logical dependency edges point to the same successor rule, the logical deduction results of each edge are weighted and fused according to the edge weights. The weights reflect the credibility of the logical dependency path, and the weighted fusion result is used as the comprehensive logical dependency satisfaction.

[0080] For the verification of timing-triggered relationships, the trigger timestamps of each rule node in the initial rule node set are extracted. These timestamps record the system time corresponding to when the rule is determined to be triggered. For a candidate successor rule node, the trigger timestamps of all predecessor rule nodes connected to it through timing-triggered relationship edges are obtained. Timing constraint parameters, including the minimum allowed time interval, are extracted from the timing-triggered relationship edges. With the maximum time interval Calculate the time difference between the assumed trigger time of the candidate successor rule node and the actual trigger time of each predecessor rule node, and determine whether this time difference falls within the time frame. Within the interval. If the time difference of at least one predecessor rule node satisfies the interval constraint, the result of the time-series trigger satisfaction evaluation is "satisfied"; otherwise, it is "not satisfied". For the case of multiple time-series trigger relationship edges, the time-series constraint satisfaction of each edge is calculated and summarized using a logical OR operation. That is, as long as any path satisfies the time-series constraint, the time-series trigger condition is determined to be satisfied.

[0081] When jointly determining the logical dependency satisfaction evaluation results and the temporal triggering satisfaction evaluation results, a logical AND operation rule is used. This means that a candidate successor rule node must simultaneously satisfy both the logical dependency condition and the temporal triggering condition to be determined as a cascaded trigger node. The logical dependency condition ensures the causal logic rationality of the rule triggering, while the temporal triggering condition ensures the temporal sequence rationality of the rule triggering; both are indispensable. For certain special rule types, where only a single type of constraint relationship exists, the satisfaction evaluation result of the missing type is defaulted to "satisfied" to avoid rule chain interruption due to missing constraints. After filtering out candidate successor rule nodes that satisfy the joint determination conditions, these nodes are marked as cascaded trigger nodes, and their connection relationships with predecessor rule nodes and corresponding rule relationship edges are recorded.

[0082] The cascaded trigger nodes identified in this round are used as the new initial rule node set to begin the next round of iterative expansion. Semantic neighborhood range calculation is repeatedly performed on each rule node in the new set. Since the semantic features of the new nodes differ from the original initial rule nodes, their semantic neighborhood range will change, allowing for the exploration of deeper rule nodes in the rule knowledge graph. Candidate successor rule nodes are identified within the new semantic neighborhood, new rule relationship edges are extracted, and constraints are resolved. Logical dependency deduction and temporal trigger verification are performed. During the iteration process, the set of visited rule nodes is maintained to avoid redundant processing of the same rule node, which could lead to computational redundancy or circular dependencies. When no new candidate successor rule nodes satisfying the joint decision conditions are identified within the semantic neighborhood of all new initial rule nodes in a given iteration, the rule chain expansion reaches its endpoint, and the iteration process terminates.

[0083] A directed graph is constructed to represent the complete rule chain by summarizing all cascaded trigger nodes identified during the iterations and their connected rule relationship edges. The initial set of rule nodes is used as the starting node of the directed graph, and the cascaded trigger nodes identified in each iteration are added to the directed graph in the order of identification. Rule relationship edges are represented as directed edges in the directed graph according to their propagation direction, pointing from the predecessor rule node to the successor rule node. The edges carry attributes such as relationship type label, constraint parameters, and satisfaction evaluation value. For branching structures in the rule chain, where a predecessor rule node triggers multiple successor rule nodes simultaneously, the directed graph represents this node as having multiple outgoing edges, each pointing to a different successor node, preserving the integrity of the branching structure. For converging structures in the rule chain, where multiple predecessor rule nodes jointly trigger the same successor rule node, the directed graph represents this successor node as having multiple incoming edges. The predecessor node and constraint conditions corresponding to each incoming edge are recorded for subsequent analysis of multi-source factors triggering the rules. The completed directed graph structure is a complete rule chain, clearly showing the propagation path and logical dependencies from the initial triggering rule to all cascading triggering rules, providing structured support for the causal explanation and rule conflict detection of subsequent review decisions.

[0084] A second aspect of this invention provides an image content compliance review system based on AI reasoning, comprising: The scene semantic unit is used to acquire the image to be reviewed and its associated multi-source context data, perform multi-granularity region segmentation and object relationship modeling on the image to be reviewed, extract the spatial topological relationship and semantic dependency relationship between object entities in the image, generate a structured scene graph representation, and perform context sensitivity weighting on the nodes in the structured scene graph representation based on the multi-source context data to form a context-aware scene semantic representation. The compliance assessment unit is used to input the scene semantic representation into the reasoning module, and perform explicit violation detection and implicit risk prediction in parallel based on the spatial topological relationship and semantic dependency relationship in the scene semantic representation to generate a multi-dimensional compliance assessment result. The decision reasoning unit is used to establish a dynamic rule knowledge graph, perform reasoning path search in the dynamic rule knowledge graph based on the multi-dimensional compliance assessment results, identify the triggered rule chains and conflict relationships between rules, generate hierarchical review decisions, and extract the rule nodes in the reasoning path and the corresponding evidence nodes in the scene semantic representation to construct the causal explanation chain of the review decision. The dynamic update unit is used to execute the hierarchical review decision, collect manual review feedback and user appeal data, use the manual review feedback and user appeal data to incrementally update the rule entity relationships in the dynamic rule knowledge graph, and at the same time adjust the context sensitivity weight in the scene semantic representation according to the causal explanation chain.

[0085] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0086] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0087] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image content compliance review method based on AI reasoning, characterized in that, include: The process involves acquiring images to be reviewed and their associated multi-source context data, performing multi-granularity region segmentation and object relationship modeling on the images to be reviewed, extracting spatial topological relationships and semantic dependencies between object entities in the images, generating a structured scene graph representation, and applying context-sensitive weighting to the nodes in the structured scene graph representation based on the multi-source context data to form a context-aware scene semantic representation. The scene semantic representation is input into the reasoning module, and explicit violation detection and implicit risk prediction are performed in parallel based on the spatial topological relationship and semantic dependency relationship in the scene semantic representation to generate multi-dimensional compliance assessment results. A dynamic rule knowledge graph is established. Based on the multi-dimensional compliance assessment results, reasoning path search is performed in the dynamic rule knowledge graph to identify the triggered rule chains and conflict relationships between rules, generate hierarchical review decisions, and extract the rule nodes in the reasoning path and the corresponding evidence nodes in the scene semantic representation to construct a causal explanation chain for the review decision. The hierarchical review decision is executed, and manual review feedback and user appeal data are collected. The manual review feedback and user appeal data are used to incrementally update the rule entity relationships in the dynamic rule knowledge graph. At the same time, the context sensitivity weight in the scene semantic representation is adjusted backtrackingly according to the causal explanation chain.

2. The method according to claim 1, characterized in that, The images to be reviewed are subjected to multi-granularity region segmentation and object relationship modeling. Spatial topological relationships and semantic dependencies between objects in the images are extracted, and a structured scene graph representation is generated, including: A multi-scale feature pyramid is constructed, and the image to be reviewed is segmented into hierarchical regions using the multi-scale feature pyramid. Candidate region sets are generated at different scale levels, and cross-level consistency checks are performed on the candidate regions at each scale level to select a stable region set with scale invariance. Object entity identification and attribute labeling are performed on each region in the stable region set to obtain an object entity set containing object category, location coordinates and semantic attributes. The relative orientation and spatial distance between objects are calculated based on the location coordinates of each object entity in the object entity set to determine the spatial topological relationship. The semantic attributes of each object entity in the object entity set are mapped to a pre-constructed semantic knowledge space. The semantic distance and co-occurrence probability between object entities in the semantic knowledge space are calculated. The semantic distance is then spatially constrained and corrected in conjunction with the spatial topological relationship to generate a relation weight matrix representing semantic dependencies. A structured scene graph representation is constructed using the object entities in the object entity set as nodes and the spatial topological relationship and the relationship weight matrix as the attributes of the edges.

3. The method according to claim 1, characterized in that, A multi-scale feature pyramid is constructed, and the image to be reviewed is segmented into hierarchical regions using the multi-scale feature pyramid. Candidate region sets are generated at different scale levels, and cross-level consistency checks are performed on the candidate regions at each scale level. The set of stable regions with scale invariance is selected, including: The images to be reviewed are downsampled step by step to generate multiple image layers with decreasing resolution. Semantic features and structural features are extracted from each image layer. A bidirectional feature transfer path is established between adjacent image layers. The semantic features and structural features of different image layers are fused across layers through the bidirectional feature transfer path to construct a multi-scale feature pyramid. Region extraction operations are performed at each image level of the multi-scale feature pyramid to generate a candidate region set corresponding to each image level. Cross-level spatial alignment is performed on the candidate region set, projecting candidate regions from different image levels onto a unified reference coordinate system. Cross-level candidate region pairs with overlapping spatial positions in the unified reference coordinate system are identified. The feature response consistency of each cross-level candidate region pair is calculated. When the feature response consistency meets a preset stability condition, the corresponding cross-level candidate region pairs are merged into a single stable region. The single stable regions are then summarized to form a stable region set.

4. The method according to claim 1, characterized in that, Based on the multi-source context data, the nodes in the structured scene graph representation are weighted by context sensitivity to form a context-aware scene semantic representation, including: Time series features, geographic location features, and user interaction features are extracted from multi-source context data, and cross-modal semantic alignment is performed on the time series features, geographic location features, and user interaction features to generate a fused context semantic embedding vector. The semantic representation of each node in the structured scene graph representation is matched with the context semantic embedding vector for similarity, and the semantic activation strength of each node under the current context condition is calculated. Based on the semantic activation intensity, a node importance ranking sequence is constructed, and a node sensitivity weight is calculated according to the relative position of each node in the node importance ranking sequence. The node sensitivity weight is then applied to the feature representation of the corresponding node in the structured scene graph representation to generate a context-weighted node feature representation. The node feature representations and their associated edge relationships are aggregated by graph convolution to form a context-aware scene semantic representation.

5. The method according to claim 1, characterized in that, The scene semantic representation is input into the inference module. Based on the spatial topological relationships and semantic dependencies in the scene semantic representation, explicit violation detection and implicit risk prediction are performed in parallel to generate multi-dimensional compliance assessment results, including: Extract object location distribution features corresponding to spatial topological relationships and object interaction pattern features corresponding to semantic dependencies from scene semantic representation, and input the object location distribution features into the explicit violation detection branch and the object interaction pattern features into the implicit risk prediction branch. In the explicit violation detection branch, the location distribution features of the object are matched with rules based on a pre-built violation rule base to identify explicit violation objects that violate explicit prohibitions and generate explicit violation detection results; In the implicit risk prediction branch, a behavioral association graph between objects is constructed based on the object interaction pattern characteristics. By analyzing the abnormal association patterns in the behavioral association graph, objects with potential hidden risks that could lead to violations are identified, and implicit risk prediction results are generated. The explicit violation detection results are integrated with the implicit risk prediction results, and the explicit violation objects and the implicit risk objects are respectively assigned violation severity scores and risk tendency scores to generate a multi-dimensional compliance assessment result that includes the violation severity scores and the risk tendency scores.

6. The method according to claim 1, characterized in that, A dynamic rule knowledge graph is established. Based on the multi-dimensional compliance assessment results, reasoning path searches are performed within the dynamic rule knowledge graph to identify triggered rule chains and conflict relationships between rules, generating tiered review decisions, including: A dynamic rule knowledge graph containing rule nodes and rule relationship edges is constructed, wherein the rule relationship edges represent the logical dependencies and temporal triggering relationships between rules, and the dynamic rule knowledge graph is updated in real time to reflect the dynamic changes of the rule base. Using the severity score and risk propensity score from the multi-dimensional compliance assessment results as query conditions, the initial set of rule nodes that match the severity score and risk propensity score are located in the dynamic rule knowledge graph. Starting from the initial set of rule nodes, a graph traversal expansion is performed along the rule relationship edges to trace the rule propagation path formed by the logical dependency relationship and the temporal triggering relationship, identify all rule nodes that are cascaded and trigger, and construct a complete rule chain.

7. The method according to claim 1, characterized in that, Starting from the initial set of rule nodes, a graph traversal is performed along the rule relationship edges to expand the graph, tracing the rule propagation path formed by the logical dependencies and the temporal triggering relationships, identifying all cascaded triggering rule nodes, and constructing a complete rule chain, including: Extract the rule semantic features and trigger condition expressions of each rule node in the initial rule node set, and calculate the semantic neighborhood range of each rule node in the dynamic rule knowledge graph based on the rule semantic features; Within the semantic neighborhood, candidate successor rule nodes connected to the initial rule node set through rule relationship edges are identified, and the precondition constraints corresponding to logical dependencies and the time sequence constraints corresponding to time sequence triggering relationships are parsed from the rule relationship edges. The state of the trigger condition expression of each rule node in the initial rule node set is used as the precondition input. Logical deduction is performed on the precondition constraint to generate a logical dependency satisfaction evaluation result. At the same time, the time sequence constraint is time sequence verified based on the trigger timestamp of each rule node in the initial rule node set to generate a time sequence trigger satisfaction evaluation result. The logical dependency satisfaction evaluation result and the timing trigger satisfaction evaluation result are jointly determined, and candidate successor rule nodes that meet the joint determination conditions are selected as cascaded trigger nodes. The cascaded triggering node is used as a new initial rule node set, and the semantic neighborhood range calculation and the candidate successor rule node identification are performed iteratively until no new cascaded triggering node is added. Summarize all the cascaded trigger nodes identified during the iteration process and the rule relationship edges they connect. Construct a directed graph structure based on the propagation direction of the rule relationship edges to form a complete rule chain.

8. An image content compliance review system based on AI reasoning, used to implement the method as described in any one of claims 1-7, characterized in that, include: The scene semantic unit is used to acquire the image to be reviewed and its associated multi-source context data, perform multi-granularity region segmentation and object relationship modeling on the image to be reviewed, extract the spatial topological relationship and semantic dependency relationship between object entities in the image, generate a structured scene graph representation, and perform context sensitivity weighting on the nodes in the structured scene graph representation based on the multi-source context data to form a context-aware scene semantic representation. The compliance assessment unit is used to input the scene semantic representation into the reasoning module, and perform explicit violation detection and implicit risk prediction in parallel based on the spatial topological relationship and semantic dependency relationship in the scene semantic representation to generate multi-dimensional compliance assessment results. The decision reasoning unit is used to establish a dynamic rule knowledge graph, perform reasoning path search in the dynamic rule knowledge graph based on the multi-dimensional compliance assessment results, identify the triggered rule chains and conflict relationships between rules, generate hierarchical review decisions, and extract the rule nodes in the reasoning path and the corresponding evidence nodes in the scene semantic representation to construct the causal explanation chain of the review decision. The dynamic update unit is used to execute the hierarchical review decision, collect manual review feedback and user appeal data, use the manual review feedback and user appeal data to incrementally update the rule entity relationships in the dynamic rule knowledge graph, and at the same time adjust the context sensitivity weight in the scene semantic representation according to the causal explanation chain.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images