An artificial intelligence-based loss settlement processing method, device, equipment and medium

Abstract

The present application relates to the technical field of artificial intelligence, and in particular to a loss determination processing method and device based on artificial intelligence, equipment and medium, which can be applied to the financial insurance scene, through acquiring multiple damage images of an accident vehicle under different shooting angles, preprocessing and image quality inspection on the multiple damage images, obtaining multiple target damage images, inputting the multiple target damage images into a preset damage detection model for damage detection, obtaining damage detection results under different visual angles, performing multi-view fusion and consistency verification on the damage detection results under different visual angles, generating loss determination parameters of a target damage area, based on the loss determination parameters of the target damage area, using a preset loss determination rule to perform loss determination calculation, and outputting a loss determination result, thereby effectively improving the efficiency and accuracy of vehicle loss determination, reducing vehicle damage misjudgment or omission caused by human intervention, and achieving efficient and accurate evaluation of vehicle damage.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a damage assessment method, apparatus, equipment and medium based on artificial intelligence. Background Technology

[0002] In recent years, automobiles have played an increasingly important role in people's daily lives. With the significant increase in car ownership, the density of vehicles on urban roads is growing, leading to a rise in traffic accidents. Therefore, rapid damage assessment of vehicles involved in collisions has become a crucial aspect of auto insurance services.

[0003] However, existing damage assessment methods largely rely on manual on-site inspections or manual review of user-uploaded data. This approach is time-consuming, resulting in low overall efficiency and difficulty in responding promptly to customer needs. Furthermore, manual damage assessment is susceptible to subjective factors, such as the surveyor's experience and judgment, which can easily lead to misjudgments or omissions in assessing vehicle damage, thus compromising accuracy and the fairness of claims processing. Therefore, improving the efficiency and accuracy of vehicle damage assessment is a pressing technical problem that needs to be solved. Summary of the Invention

[0004] This invention provides a damage assessment method, apparatus, equipment, and medium based on artificial intelligence to solve the technical problems of low efficiency and low accuracy in vehicle damage assessment.

[0005] Firstly, an artificial intelligence-based damage assessment method is provided, including: Multiple damage images of the accident vehicle from different shooting angles are acquired, and the multiple damage images are preprocessed and image quality inspected to obtain multiple target damage images. The multiple target damage images are input into a preset damage detection model for damage detection to obtain damage detection results from different perspectives. The damage detection results include damage location information, damage type, and instance segmentation mask. The damage detection results from different perspectives are fused and their consistency verified to generate damage assessment parameters for the target damage area. Based on the damage assessment parameters of the target damaged area, damage assessment is performed using preset damage assessment rules, and the damage assessment result is output.

[0006] Secondly, an artificial intelligence-based damage assessment device is provided, comprising: The acquisition module is used to acquire multiple damage images of the accident vehicle from different shooting angles, and to preprocess and perform image quality inspection on the multiple damage images to obtain multiple target damage images. The detection module is used to input the multiple target damage images into a preset damage detection model for damage detection and obtain damage detection results from different perspectives. The damage detection results include damage location information, damage type, and instance segmentation mask. The generation module is used to perform multi-view fusion and consistency verification on the damage detection results from different perspectives, and generate damage assessment parameters for the target damage area. The calculation module is used to perform damage assessment calculations based on the damage assessment parameters of the target damage area and using preset damage assessment rules, and output the damage assessment results.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described damage assessment method.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the aforementioned damage assessment method.

[0009] In summary, this invention provides an artificial intelligence-based damage assessment method, apparatus, device, and medium. It acquires multiple damage images of an accident vehicle from different shooting angles, preprocesses and performs image quality checks on these images to obtain multiple target damage images. These target damage images are then input into a preset damage detection model for damage detection, yielding damage detection results from different perspectives. The damage detection results include damage location information, damage type, and instance segmentation mask. This invention performs multi-view fusion and consistency verification on damage detection results from different perspectives to generate damage assessment parameters for the target damage area. Based on these parameters, damage assessment is calculated using pre-defined damage assessment rules, and the damage assessment results are output. As can be seen, this application effectively improves the efficiency and accuracy of vehicle damage assessment by sequentially preprocessing, detecting damage, verifying multi-view fusion, and calculating damage images from multiple perspectives. This reduces misjudgments or omissions of vehicle damage caused by human intervention and achieves efficient and accurate assessment of vehicle damage. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1This is a schematic diagram of an application environment for a damage assessment method according to an embodiment of the present invention; Figure 2 This is a schematic flowchart of a damage assessment method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a damage assessment device according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0012] 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, not all, of the embodiments of the present invention. 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.

[0013] This invention provides an artificial intelligence-based damage assessment method, which can be applied to applications such as... Figure 1 In this application environment, the client communicates with the server via a network. The server can receive multiple damage images of the accident vehicle from different shooting angles uploaded by the user through the client, and preprocess and perform image quality inspection on the multiple damage images to obtain multiple target damage images. These target damage images are then input into a preset damage detection model for damage detection, yielding damage detection results from different perspectives. The damage detection results include damage location information, damage type, and instance segmentation mask. Multi-view fusion and consistency verification are performed on the damage detection results from different perspectives to generate damage assessment parameters for the target damage area. Based on these parameters, damage assessment is calculated using preset damage assessment rules, and the damage assessment result is output. In this invention, for complex insurance entities such as vehicle insurance, the efficiency and accuracy of vehicle damage assessment are effectively improved by sequentially preprocessing, detecting damage, performing multi-view fusion verification, and calculating damage from multi-view damage images. This reduces misjudgments or omissions of vehicle damage caused by human intervention, achieving efficient and accurate assessment of vehicle damage. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0014] To illustrate the technical solution of the present invention, the present invention will be described in detail below through specific embodiments.

[0015] See Figure 2 This is a schematic flowchart of a damage assessment method based on artificial intelligence provided in an embodiment of the present invention, as shown below. Figure 2As shown, this damage assessment method can be implemented through the following steps.

[0016] S210: Acquire multiple damage images of the accident vehicle from different shooting angles, and perform preprocessing and image quality inspection on the multiple damage images to obtain multiple target damage images.

[0017] In one implementation, it is necessary to first acquire multiple damage images of the accident vehicle from different shooting angles. These multiple damage images refer to a series of image data obtained by comprehensively and meticulously photographing various damaged parts of the accident vehicle using various imaging devices and shooting tools, such as common smartphones, portable digital cameras, or dedicated professional inspection equipment. For example, accident parties or related users can use their own mobile devices to take pictures at the scene and directly upload the images to a designated data processing server; alternatively, professional on-site surveyors can carry and operate more sophisticated professional equipment to ensure standardized and high-precision image acquisition of the damaged parts. Each damage image should contain clear and identifiable information about the corresponding shooting angle. However, due to various factors such as the shooting environment, equipment performance, and operator skill, the acquired multiple damage images may often have various quality defects and potential problems. For example, the images may exhibit uneven lighting, excessive shadows, or underexposure; the background environment may be too complex, containing a large amount of interfering information; or the images themselves may have technical problems such as insufficient resolution, blurred details, or color distortion. Therefore, in order to ensure that the images can meet the requirements of subsequent in-depth analysis and automated processing, it is necessary to perform systematic preprocessing and strict image quality inspection on these original damaged images.

[0018] For example, in the insurance application field, a user's vehicle is involved in a minor collision in a parking lot, resulting in damage to the front bumper and left door. A mobile application captures multiple images of the damage from different angles. Specifically, it captures close-up images of the front and left sides of the front bumper, as well as the left door. These images are uploaded to a damage assessment system for processing, enabling the user to quickly obtain the damage assessment results for insurance claims, thereby improving the efficiency of the claims process.

[0019] Specifically, the preprocessing stage typically includes a series of targeted image optimization operations. These include improving contrast and brightness through image enhancement techniques, implementing format standardization to ensure data consistency, performing blur detection to remove unclear images, conducting angle verification to confirm that the shooting angle conforms to specifications, and using vehicle region segmentation algorithms to accurately extract the vehicle's main body region from complex backgrounds. These steps effectively improve the overall image quality and usability. In the image quality inspection stage, the preprocessed images undergo further quality verification and anomaly detection. This includes assessing whether the images meet a series of preset quality standards such as sharpness, integrity, and angular coverage, and detecting for any serious defects or abnormal data that are difficult to repair. Only through these rigorous processing and inspection steps can a series of qualified target damage images that meet the analysis requirements be obtained, providing a reliable data foundation for subsequent damage assessment, liability determination analysis, and other processes. By employing these steps, multiple target damage images can be obtained quickly and accurately, ensuring the quality and reliability of subsequent input data and avoiding subsequent misjudgments due to image quality issues.

[0020] S220: Input the multiple target damage images into a preset damage detection model to perform damage detection and obtain damage detection results from different perspectives. The damage detection results include damage location information, damage type, and instance segmentation mask.

[0021] In one implementation, multiple target damage images are input into a pre-defined damage detection model for damage detection to obtain damage detection results from different perspectives. These damage detection results include damage location information, damage type, and instance segmentation mask. The damage location information details the precise spatial coordinates of the damaged area in the input image. This information is typically presented as the coordinates of the bounding box (e.g., the coordinates of the top left and bottom right corners) or the coordinates of the center point of the damaged area. With this precise coordinate data, we can clearly infer and determine the specific location of the damage on the actual vehicle body, such as whether it is on the left side of the front bumper or the right rear door. The damage type categorizes and defines the nature of the detected damage, clearly indicating its specific form and category. Common types include, but are not limited to, surface scratches, local dents, structural cracks, or paint peeling. This provides crucial classification criteria for subsequent damage assessment. The instance segmentation mask is a more refined pixel-level output. It uses a binary image to precisely segment and label each independent damage instance in the image at the pixel level. This mask clearly outlines the complete contour of each damaged area, providing not only shape information but, more importantly, laying a solid foundation for subsequent quantitative calculations of key damage assessment parameters such as the actual area, maximum length, and perimeter of the damage. This allows the assessment work to move from qualitative judgment to quantitative analysis.

[0022] The aforementioned damage detection model is essentially a deep learning convolutional neural network (CNN) model meticulously trained and optimized with a large amount of data. This model boasts an advanced and efficient architecture, primarily composed of the following core modules working collaboratively: First, the backbone network is responsible for extracting multi-level, multi-scale basic features from the input image; second, the neck structure employs an advanced PAN-FPN (Path Aggregation Network-Feature Pyramid Network) bidirectional feature fusion mechanism, effectively integrating feature information from different depths of the backbone network to enhance feature representation capabilities; third, the feature enhancement module introduces a multi-scale attention mechanism (MSA), which adaptively focuses on key regions related to damage in the image, suppressing interference from irrelevant backgrounds, thereby improving the model's sensitivity and discriminative power regarding damage features; finally, the detection head employs an advanced anchor-free output layer design, directly predicting the location, category, and segmentation mask of the damage, simplifying the model process and improving detection efficiency. The model's training process did not start from scratch, but rather built upon a robust foundation of a mature pre-trained object detection model (such as a model trained on a large general dataset). To accurately adapt it to the specific business scenario of insurance claims, a large-scale industry-labeled dataset was constructed for transfer learning and multi-stage fine-tuning training. This dataset covers a wealth of real-world scene samples, such as various scratches on vehicle bodies, shattered windshields, and deformed or dented bumpers for vehicle damage, and cracked walls in houses and damaged critical components of industrial equipment for property damage. The dataset's annotation format is standardized and detailed, labeling each damage with its corresponding object category, specific damage type, and precise bounding box coordinates. Through continuous transfer learning and multi-stage iterative training on this high-quality, scenario-based dataset, the model ultimately achieves highly accurate and reliable identification and output of comprehensive vehicle damage detection results, providing strong technical support for automated damage assessment. For example, in a user's vehicle damage example, the model can distinguish between dents in the front bumper and scratches on the door, providing pixel-level contours, thereby greatly improving the efficiency of subsequent vehicle damage assessment.

[0023] S230: Perform multi-view fusion and consistency verification on the damage detection results from different perspectives to generate damage assessment parameters for the target damage area.

[0024] In one implementation, damage detection results from different perspectives are fused and validated across multiple viewpoints to generate damage assessment parameters for the target damage area. Since multiple target damage images are acquired from multiple viewpoints, detection and analysis from a single viewpoint often have inherent limitations and potential biases, such as incomplete identification or misjudgment due to occlusion, lighting, or angle. Therefore, multi-view information fusion aims to integrate multi-dimensional information such as damage location information, damage type, and instance segmentation masks from different perspectives. Specifically, the coordinates of the bounding boxes of the same damage area identified from different viewpoints are averaged or their union is taken to extract a damage spatial representation that more closely approximates the actual physical extent. Subsequently, consistency validation is introduced, the core of which is to evaluate the reliability between detection results from different viewpoints. For example, by calculating the area of ​​overlap (IoU) of the damage areas detected from each viewpoint in the image space, or comparing the similarity of their segmentation masks (such as the Dice coefficient), unreliable results such as false detections, duplicate detections, or low confidence levels caused by noise, algorithm errors, or extreme differences in viewpoints can be effectively identified and filtered out. Through the dual processing of multi-view fusion and consistency verification, more accurate damage assessment parameters for the target damage area can be calculated, reducing omissions or misjudgments that may occur in manual damage assessment due to perspective limitations or lack of experience, thereby providing highly reliable data support for subsequent vehicle damage assessment and claims decisions.

[0025] S240: Based on the damage assessment parameters of the target damage area, perform damage assessment calculations using preset damage assessment rules and output the damage assessment results.

[0026] In one implementation, based on the damage assessment parameters of the target damaged area, a comprehensive damage assessment calculation is performed using pre-defined damage assessment rules, ultimately generating a detailed damage assessment result. These pre-defined damage assessment rules typically consist of a series of rigorously predefined business logics or precise lookup tables. Their core function is to systematically link specific damage parameters (such as damage area, depth, and type) with corresponding repair costs or insurance compensation standards. For example, if the damage area exceeds a certain preset critical threshold, it is determined that the entire damaged component needs to be replaced; if the damage is identified as a surface scratch, the corresponding repair cost will be precisely calculated based on the measured paint area. The final damage assessment result is usually presented to the end user or relevant insurance company in a clearly structured and complete report. This report not only clarifies the scope of the loss and the repair plan but also serves as an important basis and evidence for subsequent insurance claims processes and actual repair work. Through these steps, human judgment bias and operational time are effectively reduced, thereby ensuring the fairness of the damage assessment result and improving the efficiency and accuracy of the damage assessment process.

[0027] For example, in the field of financial insurance applications, an insurance company received a report from a car owner stating that their vehicle scraped against a pillar while reversing, resulting in damage to the left rear bumper, left rear fender, and taillight. The car owner uploaded multiple photos of the vehicle damage from different angles via a mobile app, and the system was required to automatically complete the damage detection and assessment.

[0028] In summary, this invention provides an artificial intelligence-based damage assessment method, apparatus, device, and medium. It acquires multiple damage images of an accident vehicle from different shooting angles, preprocesses and performs image quality checks on these images to obtain multiple target damage images. These target damage images are then input into a preset damage detection model for damage detection, yielding damage detection results from different perspectives. The damage detection results include damage location information, damage type, and instance segmentation mask. Multi-view fusion and consistency verification are performed on the damage detection results from different perspectives to generate damage assessment parameters for the target damage area. Based on these parameters, damage assessment calculations are performed using preset damage assessment rules, and the damage assessment results are output. Therefore, this application effectively improves the efficiency and accuracy of vehicle damage assessment by sequentially preprocessing, detecting damage, performing multi-view fusion verification, and calculating damage from multiple perspective damage images. This reduces misjudgments or omissions of vehicle damage caused by human intervention, achieving efficient and accurate assessment of vehicle damage.

[0029] In one embodiment, specifically step S210, which involves preprocessing and quality checking the multiple damaged images to obtain multiple target damaged images, includes the following steps: Image enhancement processing is performed on the multiple damaged images to obtain enhanced multiple damaged images; The enhanced multiple damaged images are subjected to format normalization processing to obtain normalized multiple damaged images; The standardized multiple damage images are subjected to blur detection, angle verification, and vehicle region segmentation to obtain multiple initial damage images; The authenticity of the multiple initial damage images is verified to obtain the authenticity verification results; If the authenticity verification result is successful, then image quality inspection is performed on the multiple initial damage images to obtain multiple target damage images.

[0030] In one implementation, multiple damaged images undergo image enhancement and format standardization. Image enhancement can be achieved through various techniques, such as adaptively adjusting the brightness, contrast, and saturation of the images to compensate for insufficient or excessive lighting during capture; or employing various denoising algorithms, such as Gaussian filtering, median filtering, or nonlocal mean denoising, to eliminate random noise in the images and improve image clarity. Format standardization can also be achieved through various techniques, converting different file formats (such as PNG, BMP, TIFF, etc.) into a common format (such as JPEG), while adjusting the image resolution, size, or color space (such as converting from CMYK to RGB) to meet the input requirements of the damage detection model. This solves the problems of poor original image quality and inconsistent formats, providing a unified and clearer input basis for subsequent processing. Next, through blur detection, angle verification, and vehicle region segmentation, the blur detection can be achieved by calculating the gradient information of the image, the high-frequency components of the Fourier transform, or the response value of the Laplacian operator. The angle verification can be achieved by training an image recognition model to identify vehicle parts or the overall posture in the image and comparing it with the preset effective shooting angle. The vehicle region segmentation can be achieved through semantic segmentation or instance segmentation techniques to accurately delineate the pixel-level outline of the vehicle. In this way, multiple clear initial damage images with suitable perspectives and focused on the vehicle damage area are accurately selected from a large number of images, effectively eliminating low-quality or irrelevant information.

[0031] Based on this, the authenticity of multiple initial damage images is verified. This verification can be achieved using image metadata analysis (such as comparing EXIF ​​information with the shooting device, time, GPS location, etc.) and deepfake detection technology (identifying traces such as image stitching, copy-pasting, and inconsistent lighting). This can be accomplished through hash value comparison, EXIF ​​information detection, or cross-validation using historical vehicle data. Finally, images that pass the authenticity verification undergo rigorous image quality inspection to ensure that the resulting multiple target damage images fully meet the requirements of subsequent damage detection. These steps ensure that the subsequent damage detection model receives the highest quality input, significantly improving the accuracy of subsequent damage detection and providing a solid data foundation for subsequent damage assessment calculations. This effectively avoids assessment deviations caused by problems with the quality of the original images.

[0032] In one embodiment, specifically, performing image quality inspection on the multiple initial damage images to obtain multiple target damage images, the following steps are included: The quality of the multiple initial damage images is verified. If the multiple initial damage images pass the quality verification, it is determined whether there are any damage images belonging to a preset abnormality type among the multiple initial damage images. If there is a damaged image belonging to a preset abnormality type among the multiple initial damaged images, then the deviation of the damaged image corresponding to the preset abnormality type is calculated. The deviation is compared with a preset deviation threshold. Based on the comparison results, it is determined whether the damage image corresponding to the preset abnormality type is an abnormal image; If the damaged image corresponding to the preset abnormality type is determined to be an abnormal image, then the abnormal image is filtered to obtain multiple target damaged images.

[0033] In one implementation, multiple initial damage images undergo quality verification. This verification may include checking for severe occlusion, reflection, overexposure, or underexposure. If multiple initial damage images pass the quality verification, it is determined whether any of them belong to a preset anomaly type. This preset anomaly type refers to damage in a vehicle damage assessment scenario that is not caused by the current accident, or the image itself contains misleading or deceptive features. These types may include, but are not limited to: historical damage (not caused by the current accident), non-accident damage (such as scratches, stains, dust, and other non-structural damage), fake damage (false damage created through image editing, occlusion, etc.), and background interference unrelated to the accident (such as reflections, shadows, and occlusion by irrelevant objects). Potential anomaly features can be identified using image classification models or rule matching. Once anomaly-type damage images are identified, the system calculates their deviation. This deviation quantifies the degree of difference between the image and normal damage features. It is a quantitative indicator measuring the difference between a damage image or its local features and the preset anomaly type features. The greater the deviation, the more likely the image belongs to a certain anomaly type. For image features, the deviation can be calculated based on the statistical distance (such as Euclidean distance or Mahalanobis distance) between low-level features like texture, color, shape, and edges and anomaly-type features. For semantic features, the deviation can be calculated based on the similarity (such as cosine similarity) between high-level feature vectors extracted by deep learning models and anomaly-type feature vectors, or represented by the anomaly score output by the anomaly detection model.

[0034] Furthermore, the calculated deviation is compared with a preset deviation threshold. If the deviation exceeds the preset threshold, the image is definitively identified as an anomalous image. It is worth noting that this preset deviation threshold can be set through statistical analysis of a large number of normal and anomalous samples, establishing an empirical threshold. For example, ROC curve analysis can be used to select the threshold corresponding to the optimal F1 score. Alternatively, machine learning methods, such as Support Vector Machines (SVM) or decision trees, can be used to automatically learn and determine the classification boundary during training; the value corresponding to this boundary is the standard value. This application does not impose any limitations on this approach. Finally, to ensure the accuracy and reliability of subsequent damage assessment processes, all images identified as anomalous are filtered, ensuring that only rigorously screened, high-quality, and realistic damage images are used as multiple target damage images in the subsequent damage detection stage. This step makes the multiple target damage images obtained based on preprocessing and image quality inspection more accurate and reliable, effectively eliminating anomalous images that may lead to misjudgment or fraud, thus laying a solid foundation for subsequent damage detection and damage assessment calculations, and significantly improving the robustness and accuracy of the entire damage assessment process.

[0035] In one embodiment, specifically in step S220, which involves inputting the multiple target damage images into a preset damage detection model to perform damage detection and obtain damage detection results from different viewpoints, the steps include: The feature extraction module in the damage detection model is used to extract features from the multiple target damage images to obtain a multi-scale feature map. The feature fusion module in the damage detection model is used to perform bidirectional feature fusion on the multi-scale feature map to obtain a fused feature map. The attention mechanism module in the damage detection model is used to enhance the features of the fused feature map, generating an enhanced feature map. The enhanced feature map is decoded using the detection head in the damage detection model to obtain the damage location information, damage type, and instance segmentation mask in each target damage image; Non-maximum suppression and mask fusion are performed on the damage location information, damage type, and instance segmentation mask in each target damage image to obtain damage detection results from different perspectives.

[0036] In one implementation, when multiple target damage images are input into a pre-defined damage detection model, the feature extraction module first employs a backbone network of a deep convolutional neural network (CNN), such as ResNet, VGGNet, or EfficientNet. Through multi-layer convolution and pooling operations, it gradually extracts low-level texture features, mid-level shape features, and high-level semantic features of the images and outputs multi-scale feature maps. These multi-scale feature maps contain rich information from low-level texture to high-level semantics, providing a foundation for subsequent damage recognition. Next, the feature fusion module performs bidirectional feature fusion on these multi-scale feature maps and outputs a fused feature map. This feature fusion module can adopt a Feature Pyramid Network (FPN) structure to perform feature fusion through top-down and bottom-up paths; or it can adopt more complex fusion mechanisms such as Path Aggregation Network (PANet) or Bidirectional Feature Pyramid Network (BiFPN). This bidirectional fusion mechanism can effectively integrate information from different scales, from top-down (from high-level semantic features to low-level detail features) and bottom-up (from low-level detail features to high-level semantic features), so that each scale feature has both high-resolution details and high-level semantic context, thereby enhancing the expressive power of the features, which is especially important for detecting damage of different sizes. The fused feature maps are fed into an attention mechanism module for feature enhancement. This attention mechanism module can employ a channel attention mechanism (such as SE-Net) to recalibrate feature weights by learning the dependencies between channels; or a spatial attention mechanism (such as CBAM) to enhance features by learning the importance of spatial locations; or a Transformer-based self-attention mechanism to capture long-distance dependencies. This allows the module to intelligently identify and enhance key features related to damage while suppressing background noise and irrelevant information, enabling the model to focus more on the damage area and generate more discriminative enhanced feature maps.

[0037] Furthermore, the enhanced feature maps are received and decoded by a detection head. This detection head can be an anchor-based head, such as the RPN and classification / regression heads of Faster R-CNN; or an anchor-free head, such as FCOS or CenterNet, combined with an instance segmentation branch (such as the MaskHead of Mask R-CNN). This outputs the damage location information, damage type, and pixel-level instance segmentation mask for each target damage image. Damage location information describes the location and extent of the damage in the image, usually represented by bounding box coordinates, such as the coordinates of the top-left and bottom-right corners, or the coordinates of the center point and its width and height. Damage type refers to the label used to classify the detected damage, such as "scratch," "dent," "crack," or "paint damage." The instance segmentation mask performs pixel-level precise segmentation on each detected damage instance in the image, generating a binary mask that accurately outlines the damage contour. To ensure the uniqueness and accuracy of the detection results, non-maximum suppression (NMS) and mask fusion are applied to these preliminary detection results. NMS is a post-processing technique used to eliminate redundant and overlapping bounding boxes or masks in the detection results. Specifically, it calculates the intersection-overlap ratio (IoU) between predicted boxes, sets a threshold, retains the predicted box with the highest confidence, and suppresses other predicted boxes with an IoU exceeding the threshold to eliminate redundant overlapping predictions. Mask fusion refers to further processing of instance segmentation masks that may still have slight overlap or need to be integrated after NMS to generate a final, more accurate, and complete damage instance mask. For example, voting mechanisms, weighted averaging, or morphological operations can be used to fuse overlapping masks to generate a final accurate damage instance mask. Through these steps, vehicle damage can be accurately and comprehensively identified from multiple target damage images, solving the problem of insufficient accuracy and robustness of traditional methods in complex scenes, and laying a solid foundation for the accurate generation of final damage assessment parameters.

[0038] In one embodiment, specifically in step S230, which involves multi-view fusion and consistency verification of the damage detection results from different perspectives to generate damage assessment parameters for the target damage area, the following steps are included: Based on the damage location information, damage type and instance segmentation mask in the damage detection results, the spatial location features and appearance features of the damage area in the damage detection results from different perspectives are extracted. Based on the spatial location and appearance features of the damaged area in the damage detection results from different perspectives, the damage area from different perspectives is matched using a preset cross-view matching rule to determine a set of multi-view damaged areas belonging to the same damaged entity. A consistency analysis is performed on the set of multi-view damage regions, and target damage regions with a consistency score lower than a preset score threshold are removed based on the analysis results. Spatial fusion is performed on the instance segmentation masks in the remaining multi-view damage region set to generate the fused contour of the target damage region; Based on the fused contour, damage assessment parameters for the target damaged area are calculated, wherein the damage assessment parameters include at least one of damage area, damage length, or damage depth.

[0039] In one implementation, based on the damage location information, damage type, and instance segmentation mask in the damage detection results, spatial location features and appearance features of the damaged area are extracted from the damage detection results under different viewpoints. Spatial location features may include the center coordinates, bounding box coordinates, pixel-level contour information of the damaged area, or three-dimensional coordinates reconstructed through stereo vision technology, providing a basis for locating the damage in physical space. Appearance features may include the texture features, color features, and shape features of the damaged area, helping to distinguish different types of damage and identify the same damaged entity from different viewpoints. Subsequently, using preset cross-view matching rules, these extracted features are compared and analyzed to associate regions from different viewpoints that actually belong to the same physical damage, forming a "multi-view damage region set". Specifically, based on the camera parameters corresponding to the damage images from each viewpoint, the damage location information in each damage detection result is mapped to a three-dimensional spatial coordinate system to obtain the three-dimensional spatial coordinates of each damage region. Then, the Euclidean distance between the three-dimensional spatial coordinates of the damage regions under different viewpoints and the cosine similarity of the mask feature vectors corresponding to the instance segmentation mask are calculated. If the damage regions under any two viewpoints satisfy that the Euclidean distance is less than a preset distance threshold and the cosine similarity is greater than a preset similarity threshold, then the damage regions under the two viewpoints are determined to belong to the same damage entity and are included in the same multi-view damage region set.

[0040] Furthermore, to ensure the accuracy of damage assessment, a consistency analysis is performed on this multi-view damage region set. This assesses whether there are contradictions or uncertainties in the detection results of the same damage from different perspectives, and based on the analysis results, target damage regions with scores below a preset threshold are removed, effectively filtering out low-quality or inaccurate detection results. After removing inconsistent regions, for the remaining multi-view damage region set that has passed consistency verification, the system performs spatial fusion on the instance segmentation masks it contains to generate a more complete and accurate "fusion contour." This can be achieved using methods such as projection fusion based on 3D reconstruction, polygon intersection and union operations, or fusion based on probabilistic maps, to eliminate occlusion, deformation, or detection errors caused by different perspectives. This yields the most accurate two-dimensional or three-dimensional representation of the damage region on the vehicle surface, enabling the contour to more accurately reflect the true shape and extent of the damage region on the vehicle surface. Ultimately, based on this fused profile, the system can calculate the damage assessment parameters for the target damaged area. These parameters include at least one of damage area, damage length, or damage depth. Damage area refers to the two-dimensional projected area of ​​the damaged area on the vehicle surface; damage length refers to the maximum straight-line distance or perimeter of the damaged area; and damage depth refers to the degree of concavity or convexity of the damaged area relative to the original vehicle surface. For example, the damage area can be calculated by acquiring the projected area of ​​the fused profile on the three-dimensional surface model of the target to be assessed, based on the number of mesh patches in the projected area and the actual physical area of ​​a single mesh patch. The corresponding three-dimensional point cloud data is then extracted, and the maximum span of the three-dimensional point cloud data in the normal direction is calculated as the damage depth. Based on the damage area and / or damage depth, the damage assessment parameters for the target damaged area are determined. This step effectively integrates multi-view information, overcomes the limitations of single-view detection, significantly improves the accuracy and consistency of damage detection results, avoids damage assessment deviations caused by erroneous or low-quality detection results, and makes the final damage assessment results more objective and reliable. This improves the accuracy and efficiency of the entire damage assessment process, reduces manual intervention and errors, and enhances user satisfaction.

[0041] In one embodiment, namely, performing consistency analysis on the set of multi-view damage regions and removing target damage regions with consistency scores below a preset score threshold based on the analysis results, the following steps are included: Obtain the statistical number of viewpoints in the set of multi-view damage regions, and calculate the mean damage confidence score corresponding to the set of multi-view damage regions; Based on the statistical number of viewpoints and the mean of damage confidence, a multi-view consistency score is calculated for the multi-view damage region set. Determine whether the multi-view consistency score is lower than a preset score threshold; If the multi-view consistency score is lower than the preset score threshold, the target damage region corresponding to the multi-view damage region set is removed.

[0042] In one implementation, for a set of multi-view damage regions belonging to the same damaged entity, the comprehensiveness and average reliability of the observed damage entity are quantified by obtaining the statistical number of views and calculating the average damage confidence score corresponding to each damage region. The statistical number of views reflects how many different views the damage entity was successfully detected. This can be obtained by counting the number of image frames matching the damage entity, or by accumulating the original image source view numbers for each damage region instance. The average damage confidence score reflects the average degree of certainty of the model for these detection results. It can be obtained by weighted averaging the classification confidence, location confidence, or segmentation confidence of the damage regions detected at each view, or by directly calculating the arithmetic mean of all relevant confidence scores to obtain the average damage confidence score corresponding to the set of multi-view damage regions. Subsequently, based on these two key indicators, the system comprehensively calculates a multi-view consistency score for the multi-view damage region set. This score comprehensively considers the visibility of the damaged entity from different viewpoints, detection stability, and the model's confidence in its recognition. For example, a scoring function can be designed that takes the number of viewpoints and the mean damage confidence as input, and combines other factors (such as the overlap of damage regions and shape similarity from different viewpoints) for weighted calculation to generate a score between 0 and 1. Alternatively, fuzzy logic or expert system rules can be used to map different combinations of the number of viewpoints and the mean damage confidence to a consistency score level.

[0043] Furthermore, by comparing the calculated multi-view consistency score with a preset scoring threshold, the system can intelligently determine the reliability of the damage region set. For example, if the preset scoring threshold is 0.75 and the calculated multi-view consistency score is 0.8, the multi-view damage region set is considered reliable and does not need to be removed; otherwise, it is removed. It should be noted that the preset scoring threshold can be optimized based on historical data, expert experience, or machine learning methods such as cross-validation to balance the accuracy and recall of damage assessment. The threshold can also be dynamically adjusted according to different damage types or vehicle parts to adapt to different damage assessment scenarios; this application does not impose any limitations on this. If the consistency score is lower than the preset threshold, it indicates that the damage region set may have significant uncertainty or detection errors. For example, it may be unclear or have low confidence in some views, or be detected only in a few views, resulting in insufficient overall reliability. In this case, the system will actively remove the target damage region corresponding to the multi-view damage region set, preventing it from participating in subsequent instance segmentation mask space fusion and damage assessment parameter calculation. This method effectively filters out low-quality, inconsistent, or unreliable damage detection results, thereby significantly improving the accuracy and reliability of the final damage assessment parameters, avoiding damage assessment deviations caused by uncertain information, and thus improving the accuracy of the entire damage assessment process.

[0044] In one embodiment, after step S240, i.e. after the damage assessment result is output, the following steps are included: Determine whether the target damage area detected by the damage assessment results matches the accident type; If the target damage area detected by the damage assessment result matches the accident type, then based on the target damage area detected by the damage assessment result, a spatial logic verification of the damage location is performed to obtain the spatial logic verification result. The damage assessment results are compared with the historical damage assessment records of the accident vehicle to obtain the data comparison results; Based on the spatial logic verification results and the data comparison results, a fraud risk score is calculated; Based on the fraud risk score, a corresponding risk warning or loss assessment review request is generated.

[0045] In one implementation, the detected target damage area is first matched with the accident type to preliminarily determine the logical plausibility of the damage and the accident cause. For example, for a rear-end collision, the main damage area should be concentrated at the rear of the vehicle; for a side collision, the damage should be on the side of the vehicle. Alternatively, machine learning models can be used to learn the correspondence between different accident types and typical damage area distributions by training on a large amount of historical accident data, thereby determining the degree of matching between the current damage assessment result and the accident type. If the match is successful, spatial logic verification is further performed on the damage location to obtain spatial logic verification results, ensuring the physical possibility and consistency of the damage on the vehicle structure. For example, if damage to a door is detected, it should be verified whether there are any anomalies in the connection area between the door and the body frame, or whether the damage to adjacent components (such as the fender or B-pillar) conforms to the physical collision propagation path. Verification can also be performed using preset damage propagation path rules or component association rules. For example, if severe damage to the front bumper is detected, but adjacent components such as the headlights and radiator are intact, there may be a spatial logic inconsistency.

[0046] Furthermore, the current damage assessment results are compared with the historical damage assessment records of the accident vehicle to obtain data comparison results, thereby identifying potential duplicate claims or abnormal damage patterns. For example, it checks whether there are repeated damages to the same area within a short period of time, or whether the degree of damage is abnormal compared to historical records. Data mining techniques can also be used to analyze damage patterns and repair frequencies in historical damage assessment records to create a "damage profile" of the vehicle. The current damage assessment results are then compared with this profile to identify abnormal situations that deviate from the normal pattern. By comprehensively considering the spatial logic verification results and data comparison results, a quantitative fraud risk score can be calculated by using weighted summation or a machine learning classification model. Finally, based on this fraud risk score, the system can intelligently generate corresponding risk warnings or damage assessment review requests. For example, when the score exceeds the high-risk threshold, the system automatically generates a "damage assessment review request" with a detailed risk analysis report, which is submitted to human experts for secondary review; when the score is in the medium-risk range, a "risk warning" is generated to remind the damage assessor to pay attention to specific suspicious points. The level of detail of the warning and the priority of the review can also be dynamically adjusted based on the continuous values ​​of the risk score. This step allows for a comprehensive and multi-dimensional verification of the rationality and authenticity of the damage assessment results, effectively avoiding the risks that may arise from relying solely on physical damage assessments. It significantly improves the accuracy and reliability of the damage assessment results, thereby preventing economic losses caused by information mismatch or malicious behavior, protecting the interests of insurance companies, and enhancing the intelligence and risk control capabilities of the entire damage assessment process.

[0047] For example, in the financial insurance sector, after obtaining the damage assessment results for a motor vehicle accident, it is determined whether the frontal damage area detected in the assessment results matches the rear-end collision type submitted by the user. If they match, a high fraud risk score is calculated based on the abnormal results of spatial logic verification and frequent repetitions in historical records. If the score exceeds a preset threshold of 0.8, a high-risk damage assessment review request will be automatically generated, along with a detailed verification report and historical comparison data, and submitted to a human damage assessment expert for secondary review, while a risk warning is issued.

[0048] In one embodiment, an AI-based damage assessment device is provided, which corresponds one-to-one with the AI-based damage assessment method described in the above embodiments. For example... Figure 3 As shown, the damage assessment processing device 30 includes an acquisition module 31, a detection module 32, a generation module 33, and a calculation module 34. Detailed descriptions of each functional module are as follows: The acquisition module 31 is used to acquire multiple damage images of the accident vehicle from different shooting angles, and to preprocess and inspect the multiple damage images to obtain multiple target damage images. The detection module 32 is used to input the multiple target damage images into a preset damage detection model for damage detection and obtain damage detection results from different perspectives. The damage detection results include damage location information, damage type and instance segmentation mask. The generation module 33 is used to perform multi-view fusion and consistency verification on the damage detection results from different perspectives, and generate damage assessment parameters for the target damage area. The calculation module 34 is used to perform damage assessment calculation based on the damage assessment parameters of the target damage area and using preset damage assessment rules, and output the damage assessment result.

[0049] In one embodiment, the acquisition module 31 described above is specifically used for: Image enhancement processing is performed on the multiple damaged images to obtain enhanced multiple damaged images; The enhanced multiple damaged images are subjected to format normalization processing to obtain normalized multiple damaged images; The standardized multiple damage images are subjected to blur detection, angle verification, and vehicle region segmentation to obtain multiple initial damage images; The authenticity of the multiple initial damage images is verified to obtain the authenticity verification results; If the authenticity verification result is successful, then image quality inspection is performed on the multiple initial damage images to obtain multiple target damage images.

[0050] In one embodiment, the acquisition module 31 is further configured to: The quality of the multiple initial damage images is verified. If the multiple initial damage images pass the quality verification, it is determined whether there are any damage images belonging to a preset abnormality type among the multiple initial damage images. If there is a damaged image belonging to a preset abnormality type among the multiple initial damaged images, then the deviation of the damaged image corresponding to the preset abnormality type is calculated. The deviation is compared with a preset deviation threshold. Based on the comparison results, it is determined whether the damage image corresponding to the preset abnormality type is an abnormal image; If the damaged image corresponding to the preset abnormality type is determined to be an abnormal image, then the abnormal image is filtered to obtain multiple target damaged images.

[0051] In one embodiment, the detection module 32 described above is specifically used for: The feature extraction module in the damage detection model is used to extract features from the multiple target damage images to obtain a multi-scale feature map. The feature fusion module in the damage detection model is used to perform bidirectional feature fusion on the multi-scale feature map to obtain a fused feature map. The attention mechanism module in the damage detection model is used to enhance the features of the fused feature map, generating an enhanced feature map. The enhanced feature map is decoded using the detection head in the damage detection model to obtain the damage location information, damage type, and instance segmentation mask in each target damage image; Non-maximum suppression and mask fusion are performed on the damage location information, damage type, and instance segmentation mask in each target damage image to obtain damage detection results from different perspectives.

[0052] In one embodiment, the above-mentioned generation module 33 is specifically used for: Based on the damage location information, damage type and instance segmentation mask in the damage detection results, the spatial location features and appearance features of the damage area in the damage detection results from different perspectives are extracted. Based on the spatial location and appearance features of the damaged area in the damage detection results from different perspectives, the damage area from different perspectives is matched using a preset cross-view matching rule to determine a set of multi-view damaged areas belonging to the same damaged entity. A consistency analysis is performed on the set of multi-view damage regions, and target damage regions with a consistency score lower than a preset score threshold are removed based on the analysis results. Spatial fusion is performed on the instance segmentation masks in the remaining multi-view damage region set to generate the fused contour of the target damage region; Based on the fused contour, damage assessment parameters for the target damaged area are calculated, wherein the damage assessment parameters include at least one of damage area, damage length, or damage depth.

[0053] In one embodiment, the generation module 33 is further configured to: Obtain the statistical number of viewpoints in the set of multi-view damage regions, and calculate the mean damage confidence score corresponding to the set of multi-view damage regions; Based on the statistical number of viewpoints and the mean of damage confidence, a multi-view consistency score is calculated for the multi-view damage region set. Determine whether the multi-view consistency score is lower than a preset score threshold; If the multi-view consistency score is lower than the preset score threshold, the target damage region corresponding to the multi-view damage region set is removed.

[0054] In one embodiment, after the aforementioned calculation module 34, the following is specifically used for: Determine whether the target damage area detected by the damage assessment results matches the accident type; If the target damage area detected by the damage assessment result matches the accident type, then based on the target damage area detected by the damage assessment result, a spatial logic verification of the damage location is performed to obtain the spatial logic verification result. The damage assessment results are compared with the historical damage assessment records of the accident vehicle to obtain the data comparison results; Based on the spatial logic verification results and the data comparison results, a fraud risk score is calculated; Based on the fraud risk score, a corresponding risk warning or loss assessment review request is generated.

[0055] For specific limitations regarding the damage assessment processing device, please refer to the limitations of the damage assessment processing method above. Its specific functions and technical effects are detailed in the method implementation section and will not be repeated here. Each module in the aforementioned damage assessment processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0056] Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Figure 4 As shown, the computer device of this embodiment includes: at least one processor ( Figure 4Only one is shown in the diagram), a memory, and a computer program stored in the memory and executable on at least one processor, wherein the processor executes the computer program to perform the following steps: Multiple damage images of the accident vehicle from different shooting angles are acquired, and the multiple damage images are preprocessed and image quality inspected to obtain multiple target damage images. The multiple target damage images are input into a preset damage detection model for damage detection to obtain damage detection results from different perspectives. The damage detection results include damage location information, damage type, and instance segmentation mask. The damage detection results from different perspectives are fused and their consistency verified to generate damage assessment parameters for the target damage area. Based on the damage assessment parameters of the target damaged area, damage assessment is performed using preset damage assessment rules, and the damage assessment result is output.

[0057] This computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that... Figure 3 The examples of computer devices are merely examples and do not constitute a limitation on computer devices. Computer devices may include more or fewer components than shown in the illustration, or combinations of certain components, or different components, such as network interfaces, displays, and input systems.

[0058] In one embodiment, a computer-readable storage medium is provided that, when the instructions in the computer-readable storage medium are executed by a processor in a computer device, enables the computer device to perform the steps of any embodiment of the damage assessment method disclosed in this invention, which will not be repeated here. The computer-readable storage medium may be non-volatile or volatile.

[0059] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0060] The processor referred to can be a CPU, but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0061] Memory includes readable storage media, internal memory, etc., wherein internal memory can be the RAM of a computer device, providing an environment for the operation of the operating system and computer-readable instructions stored in the readable storage media. The readable storage media can be the hard drive of the computer device, or in other embodiments, it can be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, memory can include both internal storage units and external storage devices of the computer device. Memory is used to store the operating system, cooperative applications, boot loader, data, and other programs, such as program code of computer programs. Memory can also be used to temporarily store data that has been output or will be output.

[0062] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0063] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0064] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0065] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A damage assessment method based on artificial intelligence, characterized in that, include: Multiple damage images of the accident vehicle from different shooting angles are acquired, and the multiple damage images are preprocessed and image quality inspected to obtain multiple target damage images. The multiple target damage images are input into a preset damage detection model for damage detection to obtain damage detection results from different perspectives. The damage detection results include damage location information, damage type, and instance segmentation mask. The damage detection results from different perspectives are fused and their consistency verified to generate damage assessment parameters for the target damage area. Based on the damage assessment parameters of the target damaged area, damage assessment is performed using preset damage assessment rules, and the damage assessment result is output.

2. The damage assessment method as described in claim 1, characterized in that, The process of preprocessing and image quality inspection of the multiple damaged images to obtain multiple target damaged images includes: Image enhancement processing is performed on the multiple damaged images to obtain enhanced multiple damaged images; The enhanced multiple damaged images are subjected to format normalization processing to obtain normalized multiple damaged images; The standardized multiple damage images are subjected to blur detection, angle verification, and vehicle region segmentation to obtain multiple initial damage images; The authenticity of the multiple initial damage images is verified to obtain the authenticity verification results; If the authenticity verification result is successful, then image quality inspection is performed on the multiple initial damage images to obtain multiple target damage images.

3. The damage assessment method as described in claim 2, characterized in that, The step of performing image quality inspection on the multiple initial damage images to obtain multiple target damage images includes: The quality of the multiple initial damage images is verified. If the multiple initial damage images pass the quality verification, it is determined whether there are any damage images belonging to a preset abnormality type among the multiple initial damage images. If there is a damaged image belonging to a preset abnormality type among the multiple initial damaged images, then the deviation of the damaged image corresponding to the preset abnormality type is calculated. The deviation is compared with a preset deviation threshold. Based on the comparison results, it is determined whether the damage image corresponding to the preset abnormality type is an abnormal image; If the damaged image corresponding to the preset abnormality type is determined to be an abnormal image, then the abnormal image is filtered to obtain multiple target damaged images.

4. The damage assessment method as described in claim 1, characterized in that, The step of inputting the multiple target damage images into a preset damage detection model for damage detection, and obtaining damage detection results from different viewpoints, includes: The feature extraction module in the damage detection model is used to extract features from the multiple target damage images to obtain a multi-scale feature map. The feature fusion module in the damage detection model is used to perform bidirectional feature fusion on the multi-scale feature map to obtain a fused feature map. The attention mechanism module in the damage detection model is used to enhance the features of the fused feature map, generating an enhanced feature map. The enhanced feature map is decoded using the detection head in the damage detection model to obtain the damage location information, damage type, and instance segmentation mask in each target damage image; Non-maximum suppression and mask fusion are performed on the damage location information, damage type, and instance segmentation mask in each target damage image to obtain damage detection results from different perspectives.

5. The loss assessment method as described in claim 1, characterized in that, The process of fusing and verifying the consistency of damage detection results from different perspectives to generate damage assessment parameters for the target damage area includes: Based on the damage location information, damage type and instance segmentation mask in the damage detection results, the spatial location features and appearance features of the damage area in the damage detection results from different perspectives are extracted. Based on the spatial location and appearance features of the damaged area in the damage detection results from different perspectives, the damage area from different perspectives is matched using a preset cross-view matching rule to determine a set of multi-view damaged areas belonging to the same damaged entity. A consistency analysis is performed on the set of multi-view damage regions, and target damage regions with a consistency score lower than a preset score threshold are removed based on the analysis results. Spatial fusion is performed on the instance segmentation masks in the remaining multi-view damage region set to generate the fused contour of the target damage region; Based on the fused contour, damage assessment parameters for the target damaged area are calculated, wherein the damage assessment parameters include at least one of damage area, damage length, or damage depth.

6. The loss assessment method as described in claim 5, characterized in that, The step of performing consistency analysis on the multi-view damage region set and removing target damage regions with consistency scores lower than a preset score threshold based on the analysis results includes: Obtain the statistical number of viewpoints in the set of multi-view damage regions, and calculate the mean damage confidence score corresponding to the set of multi-view damage regions; Based on the statistical number of viewpoints and the mean of damage confidence, a multi-view consistency score is calculated for the multi-view damage region set. Determine whether the multi-view consistency score is lower than a preset score threshold; If the multi-view consistency score is lower than the preset score threshold, the target damage region corresponding to the multi-view damage region set is removed.

7. The loss assessment method as described in claim 1, characterized in that, After the output of the damage assessment results, the following are included: Determine whether the target damage area detected by the damage assessment results matches the accident type; If the target damage area detected by the damage assessment result matches the accident type, then based on the target damage area detected by the damage assessment result, a spatial logic verification of the damage location is performed to obtain the spatial logic verification result. The damage assessment results are compared with the historical damage assessment records of the accident vehicle to obtain the data comparison results; Based on the spatial logic verification results and the data comparison results, a fraud risk score is calculated; Based on the fraud risk score, a corresponding risk warning or loss assessment review request is generated.

8. A damage assessment device based on artificial intelligence, characterized in that, include: The acquisition module is used to acquire multiple damage images of the accident vehicle from different shooting angles, and to preprocess and perform image quality inspection on the multiple damage images to obtain multiple target damage images. The detection module is used to input the multiple target damage images into a preset damage detection model for damage detection and obtain damage detection results from different perspectives. The damage detection results include damage location information, damage type, and instance segmentation mask. The generation module is used to perform multi-view fusion and consistency verification on the damage detection results from different perspectives, and generate damage assessment parameters for the target damage area. The calculation module is used to perform damage assessment calculations based on the damage assessment parameters of the target damage area and using preset damage assessment rules, and output the damage assessment results.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the damage assessment method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the damage assessment method as described in any one of claims 1 to 7.

Images