A vehicle damage assessment method, system, electronic device and storage medium

Abstract

The application provides a vehicle damage assessment method and system, electronic equipment and storage medium, which are applied to the technical field of computer vision and deep learning. The method comprises the following steps: acquiring accident images and vehicle identification information, and determining a corresponding component topology graph according to the identification information; extracting visual features of the images, and mapping the visual features to nodes of the topology graph to obtain initial node features; calculating dynamic weights of component connection edges based on the features, and constructing a dynamic accident topology graph reflecting mechanical conduction strength; performing cross-domain interaction between the dynamic accident topology graph and the visual features to obtain enhanced node features; predicting component damage states based on the features, and performing consistency verification on the results according to a preset physical logic constraint; and finally outputting a structured damage assessment result containing damage information and mechanical conduction logic. In this way, the physical topology structure and the mechanical conduction logic of the vehicle can be integrated into the deep learning network, and intelligent damage assessment with causal reasoning and structural consistency is realized.

Description

Technical Field

[0001] This disclosure relates to the fields of computer vision and deep learning technology, and in particular to a method, system, electronic device and storage medium for assessing vehicle damage. Background Technology

[0002] Vehicle damage assessment is the most critical and time-consuming step in the auto insurance claims process. Its core task is to accurately identify damaged components and determine the extent of damage (such as scratches, deformation, and fractures) based on multi-angle images of the vehicle taken at the accident scene, and generate a damage assessment report that includes repair plans. However, the effectiveness of existing damage assessment technologies heavily relies on a deep understanding of the vehicle's physical structure and damage logic. In real-world claims scenarios, simply identifying surface damage textures in images is far from sufficient; the key lies in inferring hidden damage to the internal structure from visual appearances and ensuring that the assessment results conform to common sense physics.

[0003] To address this issue, existing technologies, such as object detection and semantic segmentation models based on convolutional neural networks, primarily employ pattern matching strategies based on purely visual features. While these methods have made some progress in identifying obvious external damage (such as bumper scratches), they suffer from two key limitations when dealing with complex collision accidents and internal structural damage:

[0004] First, most existing methods remain at the isolated pixel-level or component-level visual perception stage. Traditional deep learning models (such as Mask R-CNN and YOLO) typically break down the vehicle into independent detection objects, ignoring the fact that the vehicle is a rigid structure with close physical connections and mechanical transmission relationships between its components. This method cannot perceive the transmission path of collision impact forces within the vehicle. For example, when the front bumper of a vehicle suffers a violent impact, the enormous impact force is often transmitted through physical connections to the internal headlight brackets or radiator frame, causing them to break or shift. However, since these internal components may only exhibit minor deformations in appearance or be obscured by external coverings, pure visual models lacking the support of "force transmission logic" are prone to missing detections and cannot deduce the causal chain of "force on component A causing damage to component B".

[0005] Secondly, existing methods generally lack structural consistency constraints based on physical logic, leading to model outputs that often contradict common sense in physics. Existing damage assessment models are typically "black box" systems, outputting only the category with the highest probability based on image texture, lacking prior knowledge of the vehicle's assembly logic. This results in models frequently generating logically mutually exclusive erroneous predictions, such as predicting severe damage to the internal crash beam while the external bumper cover is intact; or predicting severe damage requiring replacement but failing to provide a logically consistent cause. Such damage assessment results, lacking interpretability and logical consistency, fail to meet the stringent evidentiary requirements of insurance claims, often necessitating extensive manual secondary verification and severely hindering the efficiency of automated damage assessment.

[0006] Therefore, how to design a method that can go beyond simple visual texture features and integrate the vehicle's physical topology and mechanical transmission logic into a deep learning network to achieve an intelligent damage assessment method with causal reasoning capabilities and in compliance with structural consistency constraints is a technical challenge that urgently needs to be solved in the field of intelligent auto insurance damage assessment. Summary of the Invention

[0007] This disclosure provides a vehicle damage assessment method, system, electronic device, and storage medium, which solves the technical problems of existing vision-based vehicle damage assessment methods, which lack physical common sense and logical reasoning ability, resulting in missed detection of hidden damage and prediction results that violate physical laws and are inexplicable.

[0008] According to a first aspect of this disclosure, a method for assessing vehicle damage is provided, the method comprising:

[0009] Acquire images of the vehicle in an accident and its identification information;

[0010] Based on the vehicle identification information, the corresponding component topology diagram is determined;

[0011] Visual features of the accident image are extracted and mapped to corresponding nodes of the component topology diagram to obtain initial node features;

[0012] Based on the initial node characteristics, the dynamic weights of the connecting edges in the component topology graph are calculated to construct a dynamic accident topology graph; wherein, the dynamic weights characterize the mechanical transmission correlation strength of connected components in an accident;

[0013] The dynamic accident topology map is combined with the visual features through cross-domain feature interaction to obtain enhanced node features;

[0014] The damage state of each component is predicted based on the enhanced node features, and the consistency of the prediction results is verified based on the preset physical logic constraints.

[0015] The output contains structured damage assessment results that include information on damaged components and mechanical transmission logic.

[0016] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein determining the corresponding component topology diagram based on the vehicle identification information includes:

[0017] Parse the vehicle identification information to obtain vehicle model information;

[0018] Based on the vehicle model information, the corresponding standard component topology diagram is retrieved from a preset graph database, and the semantic features of each component node are initialized; wherein, the standard component topology diagram includes a set of component nodes and a set of edges representing physical connection relationships.

[0019] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein mapping the visual features to the corresponding nodes of the component topology diagram to obtain initial node features includes:

[0020] A spatial mask for each component is generated using a semantic segmentation model. Based on the spatial mask, the visual features are aggregated to the corresponding component nodes to obtain the visual features of the component nodes.

[0021] The visual features of component nodes are fused with the semantic features of the corresponding component nodes to obtain the initial node features with fused visual features.

[0022] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the calculation of the dynamic weights of the connecting edges in the component topology graph based on the initial node characteristics includes:

[0023] For a pair of connected component nodes, calculate the mechanical transmission probability between the pair of component nodes;

[0024] Based on the mechanical transmission probability, a dynamic adjacency matrix is ​​constructed; wherein, the non-zero elements of the dynamic adjacency matrix represent the dynamic weights of the corresponding connecting edges.

[0025] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the cross-domain feature interaction between the dynamic accident topology graph and the visual features to obtain enhanced node features includes:

[0026] Based on the topological connections and node features of the dynamic accident topology graph, a graph convolution operation is performed through a graph neural network to generate a graph structure-guided spatial attention graph.

[0027] The spatial attention map is weighted and fused with the visual features extracted from the accident image to obtain visual features with enhanced structural information.

[0028] The visual features enhanced by the structural information are mapped back to the nodes of the dynamic accident topology graph to generate visual information correction values.

[0029] Based on the visual information correction amount, the state of each node in the dynamic accident topology graph is updated to obtain the enhanced node features.

[0030] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the consistency verification of the prediction results based on preset physical logic constraints includes:

[0031] Obtain a predefined set of logical constraints; wherein the set of logical constraints contains causal dependency rules for component damage states;

[0032] Calculate the violation value of the prediction result to the causal dependency rules in the set of logical constraints;

[0033] Determine whether the violation value is less than the judgment threshold. If yes, the consistency check passes; otherwise, the consistency check fails.

[0034] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the output, comprising structured damage assessment results including information on damaged components and mechanical transmission logic, includes:

[0035] A list of damaged components is generated based on the prediction results. The list of damaged components includes component identification, damage type, and repair recommendations.

[0036] Based on the dynamic weights of each connecting edge in the dynamic accident topology graph, the cause of damage is inferred for each damaged component in the list of damaged components, and mechanical transmission logic information is generated.

[0037] The damaged component list is integrated with the mechanical transmission logic information to generate and output a structured damage assessment report.

[0038] According to a second aspect of this disclosure, a vehicle damage assessment system is provided. The system includes:

[0039] The acquisition module is used to acquire accident images and vehicle identification information of the vehicle;

[0040] The determination module is used to determine the corresponding component topology diagram based on the vehicle identification information;

[0041] The mapping module is used to extract the visual features of the accident image and map the visual features to the corresponding nodes of the component topology diagram to obtain the initial node features;

[0042] A construction module is used to calculate the dynamic weights of the connecting edges in the component topology graph based on the initial node features, so as to construct a dynamic accident topology graph; wherein, the dynamic weights characterize the mechanical transmission correlation strength of connected components in an accident;

[0043] The interaction module is used to perform cross-domain feature interaction between the dynamic accident topology map and the visual features to obtain enhanced node features;

[0044] The verification module is used to predict the damage state of each component based on the enhanced node features, and to perform consistency verification on the prediction results based on preset physical and logical constraints.

[0045] The output module is used to output structured damage assessment results that include information on damaged components and mechanical transmission logic.

[0046] According to a third aspect of this disclosure, an electronic device is provided. The electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described above.

[0047] According to a fourth aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the methods according to the first and / or second aspects of this disclosure.

[0048] This disclosure matches vehicle identification information with component topology diagrams, maps visual features to topology nodes to construct initial node features, and calculates the mechanical transmission correlation strength based on the initial node features to form a dynamic accident topology diagram. Then, through cross-domain interaction between the dynamic topology diagram and visual features, it achieves bidirectional enhancement of structural and visual information. Finally, it combines physical logic constraints to complete damage state prediction and consistency verification, and outputs a structured damage assessment result with mechanical transmission logic. This disclosure relies on dynamic topology reasoning, mechanical transmission correlation modeling, and physical logic constraints to effectively solve the problems of missed hidden damage, predictions that violate physical common sense, and lack of interpretability caused by traditional damage assessment relying solely on visual features. It significantly improves the comprehensiveness and accuracy of damage identification and the logical consistency of damage assessment conclusions, reduces the cost of manual review, and adapts to the needs of automation and standardization in auto insurance claims.

[0049] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0050] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of this disclosure. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0051] Figure 1 A flowchart of a vehicle damage assessment method according to an embodiment of the present disclosure is shown;

[0052] Figure 2 A block diagram of a vehicle damage assessment system according to an embodiment of the present disclosure is shown;

[0053] Figure 3 A block diagram of an exemplary electronic device capable of implementing embodiments of the present disclosure is shown. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0055] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0056] This disclosure matches vehicle identification information with component topology diagrams, maps visual features to topology nodes to construct initial node features, and calculates the mechanical transmission correlation strength based on the initial node features to form a dynamic accident topology diagram. Then, through cross-domain interaction between the dynamic topology diagram and visual features, it achieves bidirectional enhancement of structural and visual information. Finally, it combines physical logic constraints to complete damage state prediction and consistency verification, and outputs a structured damage assessment result with mechanical transmission logic. This disclosure relies on dynamic topology reasoning, mechanical transmission correlation modeling, and physical logic constraints to effectively solve the problems of missed hidden damage, predictions that violate physical common sense, and lack of interpretability caused by traditional damage assessment relying solely on visual features. It significantly improves the comprehensiveness and accuracy of damage identification and the logical consistency of damage assessment conclusions, reduces the cost of manual review, and adapts to the needs of automation and standardization in auto insurance claims.

[0057] Figure 1A flowchart of a vehicle damage assessment method 100 according to an embodiment of the present disclosure is shown. Figure 1 As shown, method 100 includes:

[0058] S101, acquire accident images of the vehicle and vehicle identification information.

[0059] In some embodiments, the system receives multi-angle image sequences of the accident scene uploaded by the user through a data interface. To ensure the standardization and efficiency of model processing, each original image needs to be processed. Perform standardized preprocessing procedures:

[0060] Size standardization: A bilinear interpolation algorithm is used to uniformly adjust the image resolution to [size value missing]. Pixels, to eliminate the computational complexity caused by the varying sizes of the original images;

[0061] Data normalization: The adjusted image is Z-score normalized to convert it into a tensor that meets the model input requirements. The specific calculation formula is as follows:

[0062] ;

[0063] in, The pixel matrix before normalization. The mean vector used during pre-training on the ImageNet dataset. This is the corresponding standard deviation vector. The final image sequence is converted to a vector with dimension [missing information]. The input tensor.

[0064] In some embodiments, the system synchronously receives the Vehicle Identification Number (VIM), which serves as the vehicle's "identity card" and contains key information such as model, year, and configuration.

[0065] S102, Based on the vehicle identification information, determine the corresponding component topology diagram.

[0066] In some embodiments, vehicle identification information is parsed to obtain vehicle model information; based on the vehicle model information, the corresponding standard component topology graph is retrieved from a preset graph database, and the semantic features of each component node are initialized; wherein, the standard component topology graph includes a set of component nodes and a set of edges representing physical connection relationships.

[0067] Specifically, the system has a pre-built VIN code parsing library. By parsing the World Manufacturer Identifier (WMI), Vehicle Description Section (VDS), and Vehicle Indicator Section (VIS) in the VIN code, the specific vehicle model can be accurately determined. Based on the parsed vehicle model information, the system retrieves the corresponding topology from a pre-built standard component topology database. , where the set of nodes Including car models Key components, edge set It is a symmetric adjacency matrix used to define the physical connection relationships between components. If the components With components If they are physically connected (such as the front bumper cover being attached to the headlight bracket via clips and screws), then Otherwise, it is 0. This matrix defines the potential transmission path of impact force on the vehicle structure.

[0068] In some embodiments, to enable nodes in the topology graph to possess semantic information that can be understood by the neural network, the system initializes a feature vector for each component node. Specifically, a pre-trained language model (such as BERT) is used to encode the name text of each component (e.g., "left front headlight"). A semantic feature vector of d=256 dimensions is generated by extracting the output of the model's tags or by performing average pooling on the word vectors. This vector contains the semantic information of the component in the general corpus, laying the foundation for subsequent fusion with visual features.

[0069] S103, extract the visual features of the accident image and map the visual features to the corresponding nodes of the component topology diagram to obtain the initial node features.

[0070] In some embodiments, a spatial mask for each component is generated using a semantic segmentation model. Based on the spatial mask, visual features are aggregated to the corresponding component nodes to obtain the visual features of the component nodes. The visual features of the component nodes are then fused with the semantic features of the corresponding component nodes to obtain the initial node features with fused visual features.

[0071] Specifically, the preprocessed image tensor The input is fed into a pre-trained convolutional neural network (such as the EfficientNet-B4 model pre-trained on the ImageNet dataset), the top global pooling layer and classification layer are removed, and the high-dimensional, dense feature map output from the last convolutional block is extracted. The dimension of this feature map is (e.g., C=1792, The feature vector at each spatial location encodes rich visual contextual information (such as edges, textures, and shape patterns) for a corresponding region in the original image, making it suitable for capturing subtle features of damage.

[0072] ;

[0073] in, For the extracted feature tensor; Number of channels; The height of the feature map; This represents the width of the feature map.

[0074] In some embodiments, to achieve accurate mapping from visual features to topological graph nodes, a parallel semantic segmentation branch is required. This branch is a deep learning model (e.g., based on the DeepLabV3+ architecture) trained on labeled data (pixel-level part labels). Its function is to perform pixel-level classification on the input image, assigning a part category label to each pixel in the image. The output of this branch is a series of part probability masks. Each channel M corresponds to a specific component, and each value in the mask represents the probability that the spatial location belongs to that component, which is equivalent to generating a detailed "component map".

[0075] In some embodiments, the mask is used to perform weighted aggregation of visual features, i.e., region of interest alignment (ROIPooling), and the calculation component is used. Aggregated visual features :

[0076] ;

[0077] in, For components Aggregated visual features; Number of images; Hadamard product (element-by-element multiplication); For the first The first one in the picture The probability mask for each component.

[0078] In some embodiments, structured semantic knowledge derived from the topological graph (initial semantic features of nodes) (and the perceptual appearance information from the image (aggregated visual features of nodes)) (to be integrated)

[0079] Specifically, through a linear projection layer From the visual feature dimension Down to This is then fused with the initial semantic features to obtain the initial node state incorporating visual information. :

[0080] ;

[0081] in, The projective weight matrix is ​​a learnable matrix. Semantic features for initialization.

[0082] S104. Based on the initial node characteristics, calculate the dynamic weights of the connecting edges in the component topology graph to construct a dynamic accident topology graph; whereby the dynamic weights characterize the mechanical transmission correlation strength of connected components in an accident.

[0083] In some embodiments, for connected component node pairs, the mechanical transmission probability between component node pairs is calculated; based on the mechanical transmission probability, a dynamic adjacency matrix is ​​constructed; wherein, the non-zero elements of the dynamic adjacency matrix represent the dynamic weights of the corresponding connecting edges.

[0084] In some embodiments, in the standard topology diagram In this context, an edge connection only indicates a physical connection between components, and a dynamic weight is calculated for each such edge connection. This weight is a probability value that quantitatively represents the correlation strength of the impact force transmitted from component i to component j (or vice versa) in a specific accident. For example, a high weight (such as...) >0.8) strongly indicates the component j The damage is most likely caused by the component i Caused by the force transmitted from it.

[0085] Specifically, the dynamic weights are calculated using a learnable neural network module (such as a bilinear interaction model), whose input is the fused features of connected component node pairs. For each existing edge in the standard graph (i.e.... =1 node pair Perform the following calculations:

[0086] (1) Feature concatenation: combine the initial node features of nodes i and j and Concatenate them and calculate their absolute difference vectors. To capture the differences between the two states;

[0087] (2) Nonlinear transformation and probability mapping: The concatenated feature vector is input into a small feedforward neural network, and the final probability value is output through the Sigmoid activation function. The calculation formula can be expressed as:

[0088] ;

[0089] in, For force transmission probability weights; Use the Sigmoid activation function; This is a tensor splicing operation; This is an absolute difference vector used to capture damage difference features; This is the weight matrix of the fully connected layer; For hidden layer dimensions; To output the projection vector; This is a bias term.

[0090] In some embodiments, based on the calculated dynamic weights of all connected edges. Construct a dynamic adjacency matrix for accidents. :

[0091] ;

[0092] in, (Typically set to 1.0) represents the node self-loop weight, used to preserve the node's own information in subsequent graph convolution operations. This yields the dynamic accident topology graph. =( V , ), its node set V The edge set remains unchanged. From the weight matrix definition.

[0093] S105, cross-domain feature interaction between dynamic accident topology map and visual features is performed to obtain enhanced node features.

[0094] In some embodiments, based on the topological connections and node features of the dynamic accident topology graph, a graph convolution operation is performed through a graph neural network to generate a graph structure-guided spatial attention graph; the spatial attention graph is weighted and fused with visual features extracted from the accident image to obtain structurally enhanced visual features; the structurally enhanced visual features are mapped back to the nodes of the dynamic accident topology graph to generate visual information correction values; based on the visual information correction values, the state of each node in the dynamic accident topology graph is updated to obtain enhanced node features.

[0095] In some embodiments, to address the deep integration of visual perception and structural reasoning, the core idea is to establish a two-way, iterative dialogue mechanism that allows macroscopic structural logic (topological graph) and microscopic visual evidence (image features) to guide and correct each other, thereby achieving more accurate and robust perception and reasoning of damage. This process is typically executed iteratively multiple times (e.g., T=3 times) to ensure full information exchange.

[0096] In some embodiments, a dynamic accident topology graph is used. The underlying mechanical transmission logic guides the visual system to focus on key areas:

[0097] Dynamic accident topology diagram When input into a Graph Convolutional Network (GCN), the GCN enables each node to aggregate features from its neighbors by propagating information across the graph, thereby generating enhanced node features that contain global structural context information. Enhanced graph node features As a query, deep visual feature map Spatial attention graphs are computed using an attention mechanism (such as scaled dot product attention) as keys and values. :

[0098] ;

[0099] in, For attention maps; For the query projection matrix; The key projection matrix; As a visual feature, This represents the scaling factor.

[0100] In some embodiments, the generated spatial attention map is used to weight the original visual features, highlighting the regions most relevant to the current structural reasoning:

[0101] ;

[0102] in, The value projection vector. This is the Hadamard product (element-by-element multiplication). Visual features that enhance structural information.

[0103] In some embodiments, the enhanced visual features are fed back into the graph structure to correct the initial judgments of nodes based on structure reasoning:

[0104] Using semantic segmentation masks Enhanced visual features The data is then aggregated again onto each corresponding component node to generate visual information correction values. visual correction amount With the current state of the node Together, they are input into the Gated Recurrent Unit (GRU) to update the node state:

[0105] ;

[0106] in, For the updated node The state vector; For the input of the GRU; This is the hidden state of the GRU. After... After rounds of iteration, the final node feature representation is obtained. .

[0107] In some embodiments, after T iterations, the node features fully integrate visual and structural information to obtain enhanced node features. These features are the final comprehensive representation of the state of each component, derived from a combination of visual appearance and physical logic.

[0108] S106, based on the enhanced node features, predict the damage state of each component, and perform consistency verification on the prediction results based on the preset physical logic constraints.

[0109] In some embodiments, a predefined set of logical constraints is obtained; wherein the set of logical constraints contains causal dependency rules for component damage states; the violation value of the prediction result against the causal dependency rules in the set of logical constraints is calculated; it is determined whether the violation value is less than the judgment threshold, and if so, the consistency check is passed; otherwise, the consistency check is not passed.

[0110] In some embodiments, to achieve refined loss assessment, the system enhances node characteristics. Three parallel prediction heads are connected, each corresponding to a different prediction task. The system's total loss function... It consists of four weighted parts:

[0111] ;

[0112] in, , and To balance the hyperparameters of the task weights, the specific definitions of each component loss function are as follows:

[0113] (1) Damage detection loss

[0114] This loss is used to supervise the model in determining whether a component is damaged (a binary classification task), and is calculated using the weighted binary cross-entropy loss (BCE Loss):

[0115] ;

[0116] in, This represents the total number of components. For components The actual damage label (1 for damaged, 0 for intact); The damage probability predicted by the model; These are the weights for positive samples, used to address the imbalance between positive and negative samples. The binary classifier outputs a probability value for each component node. ∈(0,1) represents the confidence level that the component is damaged. Typically, if... If the value is greater than 0.5, the component is considered damaged.

[0117] (2) Classification of damage types and losses

[0118] This loss is calculated only for truly damaged parts and is used to supervise the model in identifying specific damage types (such as deformation, scratches, breakage, etc.). It is calculated using multi-class cross-entropy loss.

[0119] ;

[0120] in, It is a collection of actual damaged parts; This refers to the number of damaged parts; The total number of predefined damage types; For components Belongs to the True labels for each type of damage (One-hot encoding); The model predicts that this component belongs to the first... Probability of each damage type. For each component identified as damaged, predict its specific damage type (e.g., scratches, deformation, fractures, dents, etc.) and output the probability distribution for each type.

[0121] (3) Classification of losses in maintenance plans

[0122] This loss is also calculated only for the damaged parts, and is used to supervise the model to output correct repair suggestions (such as sheet metal work, replacement, painting, etc.). The calculation formula is:

[0123] ;

[0124] in, The total number of predefined maintenance plans; For components The corresponding repair plan's authentic label; This represents the probabilities of repair options predicted by the model. For each damaged component, the optimal repair option (e.g., no repair required, repair, replacement, etc.) is recommended, and the probability distribution of each option is output.

[0125] (4) Structural consistency loss

[0126] This loss utilizes the constructed logical constraint matrix The forced model predictions conform to common sense physics, and this set is predefined based on vehicle repair manuals and expert experience. It consists of a series of causal rules in the form of "if-then," encoding the vehicle's assembly and mechanical common sense. This matrix serves as the criterion for logical consistency checks. For example:

[0127] Rule (u,v): If an internal structural component u (such as a crash beam) is damaged, then its external covering component v (such as a bumper cover) must also be damaged;

[0128] Rule (p,q): If the mounting bracket p is damaged, then the component q that it is fixed to (such as the headlight) will inevitably be damaged.

[0129] In some embodiments, the verification process is implemented by calculating a quantified structural consistency loss, Lconsist, which measures the degree to which the predictions violate physical-logical constraints. For each rule in the set of logical constraints... (i.e., "if internal components") Damage to the outer cover "Inevitably suffer damage", calculate its violation value:

[0130] ;

[0131] in, This is the tolerance threshold; and Components and The damage has a predictable probability. If the model predicts that the damage probability of internal components is significantly higher than that of external components (violating physical logic), then this term generates a positive loss, forcing the model parameters to be updated in a logically consistent direction.

[0132] In some embodiments, the sum of all rule violation values ​​is the total violation value. If the value is always violated ( For a very small positive value, such as If the deviation is positive, the prediction result is considered to have passed the consistency test; otherwise, it is considered to have failed. The total violation value is... Represented as:

[0133] ;

[0134] S107 outputs a structured damage assessment result that includes information on damaged components and mechanical transmission logic.

[0135] In some embodiments, a list of damaged components is generated based on the prediction results. The list of damaged components includes component identification, damage type and maintenance recommendations. Based on the dynamic weights of each connecting edge in the dynamic accident topology graph, the cause of damage is inferred for each damaged component in the list of damaged components, and mechanical transmission logic information is generated. The list of damaged components and the mechanical transmission logic information are integrated to generate and output a structured damage assessment report.

[0136] In some embodiments, all component nodes are traversed to filter out those with the probability of damage. For components with a damage value >0.5, a detailed list of damaged components is generated. For each damaged component in the list, the system performs the following mapping operation:

[0137] Component Identifier: Read the standard name of the node (e.g., "Front Bumper Cover") from the standard component topology diagram.

[0138] Damage type: Take the category index with the highest probability in the damage type classification header and map it to the predefined damage type text description (such as "severe deformation").

[0139] Repair suggestion: Take the category index with the highest probability in the repair solution category header and map it to the predefined repair solution text description (such as "replace").

[0140] In some embodiments, the dynamic accident topology graph and its dynamic adjacency matrix are constructed by backtracking. For each damaged component j in the list of damaged components, reason about the cause of the damage, specifically including:

[0141] In the dynamic graph, find all in-degree edges pointing to this component. weight If the maximum weight If it is the point of direct impact, then it is determined to be the point of direct force application, and the field is filled with "direct impact point"; if the maximum weight If so, it is determined that the transmission is impaired, and the corresponding source node is obtained. The name field is filled with "caused by force transmission from [part name]" (e.g., "caused by force transmission from the front bumper").

[0142] In some embodiments, the above information and consistency verification results are integrated to generate a final structured loss assessment report, which uses a standardized data format (such as JSON format).

[0143] According to the embodiments of this disclosure, the following technical effects are achieved:

[0144] (1) Possesses physical logic reasoning ability, effectively reducing the false negative rate: This disclosure constructs a dynamic accident topology map The model calculates the force correlation between nodes, simulating the real physical force transmission process. This allows the model to not only identify explicit external damage, but also to deduce the internal hidden components damaged by force transmission (such as headlight brackets), significantly improving the comprehensiveness of damage assessment.

[0145] (2) Introduce structural consistency constraints to ensure a rigorous chain of evidence: by constructing a logical constraint matrix And introduce structural consistency loss during training. This disclosure forces the model to learn the physical assembly rules of the vehicle. This effectively avoids predictions that defy common sense, such as "internal damage, external integrity," and ensures the physical and logical consistency of the damage assessment conclusion.

[0146] (3) Fine-grained cross-domain feature interaction is achieved: the spatial attention map generated by the structural information can guide the visual network to focus on key stress areas (such as tiny cracks or deformation points), realizing the deep integration of macroscopic structural logic and microscopic visual features, and improving the recognition accuracy of minor damage.

[0147] (4) Generate interpretable structured reports: Unlike traditional models that only output labels, this disclosure can generate text descriptions containing causal logic based on the edge weights of dynamic graphs, and automatically perform structural integrity checks, directly outputting structured JSON reports that conform to business standards, which greatly reduces the workload of manual writing and review.

[0148] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this disclosure is not limited to the described order of actions, because according to this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this disclosure.

[0149] The above is an introduction to the method embodiments. The following describes the solution described in this disclosure further through device embodiments.

[0150] Figure 2 A block diagram of a vehicle damage assessment system 200 according to an embodiment of the present disclosure is shown. Figure 2 As shown, the device 200 includes:

[0151] The acquisition module 201 is used to acquire accident images and vehicle identification information of the vehicle;

[0152] The determination module 202 is used to determine the corresponding component topology diagram based on the vehicle identification information;

[0153] The mapping module 203 is used to extract the visual features of the accident image and map the visual features to the corresponding nodes of the component topology diagram to obtain the initial node features;

[0154] Module 204 is used to calculate the dynamic weights of the connecting edges in the component topology graph based on the initial node features, so as to construct a dynamic accident topology graph; wherein, the dynamic weights characterize the mechanical transmission correlation strength of connected components in an accident;

[0155] Interaction module 205 is used to perform cross-domain feature interaction between the dynamic accident topology map and visual features to obtain enhanced node features;

[0156] The verification module 206 is used to predict the damage state of each component based on the enhanced node features, and to perform consistency verification on the prediction results based on preset physical and logical constraints.

[0157] Output module 207 is used to output structured damage assessment results that include information on damaged components and mechanical transmission logic.

[0158] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0159] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0160] Figure 3 A schematic block diagram of an electronic device 300 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0161] Electronic device 300 includes a computing unit 301, which can perform various appropriate actions and processes according to a computer program stored in ROM 302 or a computer program loaded into RAM 303 from storage unit 308. RAM 303 can also store various programs and data required for the operation of electronic device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via bus 304. I / O interface 305 is also connected to bus 304.

[0162] Multiple components in electronic device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of displays, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows electronic device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0163] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as method 100. For example, in some embodiments, method 100 may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by the computing unit 301, one or more steps of method 100 described above may be performed. Alternatively, in other embodiments, the computing unit 301 may be configured to perform method 100 by any other suitable means (e.g., by means of firmware).

[0164] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0165] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0166] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0167] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including voice input, speech input, or tactile input).

[0168] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0169] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0170] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0171] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for assessing vehicle damage, characterized in that, include: Acquire images of the vehicle in an accident and its identification information; Based on the vehicle identification information, the corresponding component topology diagram is determined; Visual features of the accident image are extracted and mapped to corresponding nodes of the component topology diagram to obtain initial node features; Based on the initial node characteristics, the dynamic weights of the connecting edges in the component topology graph are calculated to construct a dynamic accident topology graph; wherein, the dynamic weights characterize the mechanical transmission correlation strength of connected components in an accident; The dynamic accident topology map is combined with the visual features through cross-domain feature interaction to obtain enhanced node features; The damage state of each component is predicted based on the enhanced node features, and the consistency of the prediction results is verified based on the preset physical logic constraints. The output includes structured damage assessment results containing information on damaged components and mechanical transmission logic; The calculation of the dynamic weights of the connecting edges in the component topology graph based on the initial node features includes: For a pair of connected component nodes, calculate the mechanical transmission probability between the pair of component nodes; Based on the mechanical transmission probability, a dynamic adjacency matrix is ​​constructed; wherein, the non-zero elements of the dynamic adjacency matrix represent the dynamic weights of the corresponding connecting edges; The dynamic weights are calculated using a learnable neural network module. The input to this module is the fused features of connected component node pairs. For each existing edge in the standard graph, the following calculation is performed: (1) Feature concatenation: combine the initial node features of nodes i and j and Concatenate them and calculate their absolute difference vectors. To capture the differences between the two states; (2) Nonlinear transformation and probability mapping: The concatenated feature vector is input into a small feedforward neural network, and the final probability value is output through the Sigmoid activation function. The calculation formula can be expressed as: ； in, For force transmission probability weights; Use the Sigmoid activation function; This is a tensor splicing operation; This is an absolute difference vector used to capture damage difference features; This is the weight matrix of the fully connected layer; For hidden layer dimensions; To output the projection vector; This is a bias term.

2. The method according to claim 1, characterized in that, The step of determining the corresponding component topology diagram based on the vehicle identification information includes: Parse the vehicle identification information to obtain vehicle model information; Based on the vehicle model information, the corresponding standard component topology diagram is retrieved from a preset graph database, and the semantic features of each component node are initialized; wherein, the standard component topology diagram includes a set of component nodes and a set of edges representing physical connection relationships.

3. The method according to claim 2, characterized in that, The step of mapping the visual features to the corresponding nodes of the component topology diagram to obtain initial node features includes: A spatial mask for each component is generated using a semantic segmentation model. Based on the spatial mask, the visual features are aggregated to the corresponding component nodes to obtain the visual features of the component nodes. The visual features of component nodes are fused with the semantic features of the corresponding component nodes to obtain the initial node features with fused visual features.

4. The method according to claim 1, characterized in that, The step of performing cross-domain feature interaction between the dynamic accident topology map and the visual features to obtain enhanced node features includes: Based on the topological connections and node features of the dynamic accident topology graph, a graph convolution operation is performed through a graph neural network to generate a graph structure-guided spatial attention graph. The spatial attention map is weighted and fused with the visual features extracted from the accident image to obtain visual features with enhanced structural information. The visual features enhanced by the structural information are mapped back to the nodes of the dynamic accident topology graph to generate visual information correction values. Based on the visual information correction amount, the state of each node in the dynamic accident topology graph is updated to obtain the enhanced node features.

5. The method according to claim 1, characterized in that, The consistency verification of the prediction results based on preset physical and logical constraints includes: Obtain a predefined set of logical constraints; wherein the set of logical constraints contains causal dependency rules for component damage states; Calculate the violation value of the prediction result to the causal dependency rules in the set of logical constraints; Determine whether the violation value is less than the judgment threshold. If yes, the consistency check passes; otherwise, the consistency check fails.

6. The method according to claim 1, characterized in that, The output includes structured damage assessment results containing information on damaged components and mechanical transmission logic, including: A list of damaged components is generated based on the prediction results. The list of damaged components includes component identification, damage type, and repair recommendations. Based on the dynamic weights of each connecting edge in the dynamic accident topology graph, the cause of damage is inferred for each damaged component in the list of damaged components, and mechanical transmission logic information is generated. The damaged component list is integrated with the mechanical transmission logic information to generate and output a structured damage assessment report.

7. A vehicle damage assessment system, characterized in that, include: The acquisition module is used to acquire accident images and vehicle identification information of the vehicle; The determination module is used to determine the corresponding component topology diagram based on the vehicle identification information; The mapping module is used to extract the visual features of the accident image and map the visual features to the corresponding nodes of the component topology diagram to obtain the initial node features; A construction module is used to calculate the dynamic weights of the connecting edges in the component topology graph based on the initial node features, so as to construct a dynamic accident topology graph; wherein, the dynamic weights characterize the mechanical transmission correlation strength of connected components in an accident; The interaction module is used to perform cross-domain feature interaction between the dynamic accident topology map and the visual features to obtain enhanced node features; The verification module is used to predict the damage state of each component based on the enhanced node features, and to perform consistency verification on the prediction results based on preset physical and logical constraints. The output module is used to output structured damage assessment results that include information on damaged components and mechanical transmission logic. The calculation of the dynamic weights of the connecting edges in the component topology graph based on the initial node features includes: For a pair of connected component nodes, calculate the mechanical transmission probability between the pair of component nodes; Based on the mechanical transmission probability, a dynamic adjacency matrix is ​​constructed; wherein, the non-zero elements of the dynamic adjacency matrix represent the dynamic weights of the corresponding connecting edges; The dynamic weights are calculated using a learnable neural network module. The input to this module is the fused features of connected component node pairs. For each existing edge in the standard graph, the following calculation is performed: (1) Feature concatenation: combine the initial node features of nodes i and j and Concatenate them and calculate their absolute difference vectors. To capture the differences between the two states; (2) Nonlinear transformation and probability mapping: The concatenated feature vector is input into a small feedforward neural network, and the final probability value is output through the Sigmoid activation function. The calculation formula can be expressed as: ； in, For force transmission probability weights; Use the Sigmoid activation function; This is a tensor splicing operation; This is an absolute difference vector used to capture damage difference features; This is the weight matrix of the fully connected layer; For hidden layer dimensions; To output the projection vector; This is a bias term.

8. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method described in any one of claims 1-6.

Images