A large model-based loss determination method and system

Abstract

The application discloses a large model-based loss determination method and system, and belongs to the technical field of loss determination of accident vehicles, wherein the system comprises a data acquisition unit, a loss determination unit and a storage unit, the data acquisition unit is in communication connection with the loss determination unit and the storage unit respectively, and the loss determination unit is in communication connection with the storage unit. The application can construct a training data set through historical accident images and videos, and train a loss determination large model by using an exposure disturbance enhancement and stability constraint method, so that the loss determination large model can stably identify a damage area, a crack continuous structure and a fracture feature, and further generate a vehicle loss determination evaluation result.

Description

Technical Field

[0001] This invention relates to the field of accident vehicle damage assessment technology, specifically a damage assessment method and system based on a large model. Background Technology

[0002] With the increasing demand for vehicle accident damage assessment, traditional vehicle damage assessment mainly relies on manual review of accident images and experience to determine the damaged location, type, repair method, and cost. This approach is easily affected by differences in personnel experience, leading to inconsistent assessment results. While existing image recognition damage assessment methods can identify vehicle parts and damaged areas, vehicle accident images are often affected by factors such as shooting angle, lighting intensity, paint reflection, shadows, overexposure, underexposure, blurring, and compression noise. These factors can easily misjudge reflections, stains, or shadow edges as cracks, and may also miss actual cracks or fracture edges.

[0003] Especially for exterior components such as bumpers, fenders, doors, and headlights, cracks, fractures, scratches, assembly gaps, and light reflections have similar characteristics in images, making it difficult to reliably determine the extent of damage and repairability based on the results of a single image. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to provide a damage assessment method and system based on a large model, which can construct a training dataset through historical accident images and videos, and use exposure perturbation enhancement and stability constraints to train the large damage assessment model, so that it can stably identify the damage area, crack continuity structure and fracture characteristics, and further generate vehicle damage assessment results.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] A loss assessment method based on a large model includes:

[0007] S100) Data Acquisition: Acquire image and video data of the accident vehicle;

[0008] S200) The large-scale damage assessment model is used to identify the image and video data collected in step S100 and to assess the damage based on the identification results; wherein, the large-scale damage assessment model is trained through the following steps:

[0009] S201) Collect historical accident vehicle image data and historical accident vehicle video data;

[0010] S202) The video image frames of the historical accident vehicle video data collected in step S201 are extracted by combining fixed time interval frame sampling and motion change detection, and the extracted video image frames are filtered to obtain a set of effective video image frames of historical accident vehicles.

[0011] S203) The historical accident vehicle image data collected in step S201 and the effective video image frame set of historical accident vehicles obtained in step S202 are cleaned, deduplicated, and damaged to obtain a set of original damage samples of historical accident vehicles with vehicle damage labels. Then, exposure perturbation enhancement processing is performed on the original damage samples in the set of original damage samples of historical accident vehicles to obtain a dataset of damage samples of historical accident vehicles. The vehicle damage label is one or more of the following: damage type label, damage area label, crack area label, fracture area label, and repairability level label.

[0012] (S204) Divide the historical accident vehicle damage sample dataset obtained in step S203 into a training set, a validation set, and a test set in a ratio of 7:2:1, and use the training set, validation set, and test set to train the damage assessment model.

[0013] In the above-mentioned damage assessment method based on a large model, in step S203, the exposure disturbance enhancement is one or more of the following: overexposure enhancement, underexposure enhancement, local shadow superposition, blur disturbance, and compression disturbance, and the original vehicle damage label remains unchanged when performing exposure disturbance enhancement on the original damage sample of historical accident vehicles.

[0014] In the aforementioned large-scale model-based damage assessment method, in step S204, when testing the trained large-scale damage assessment model, the original damage samples of historical accident vehicles and the exposure-enhanced damage samples obtained after applying different exposure perturbation enhancement processes to the original damage samples are divided into a damage sample group. The trained large-scale damage assessment model is used to identify this damage sample group and calculate the identification stability evaluation value. When the identification stability evaluation value is greater than or equal to a preset evaluation threshold, it indicates that the large-scale damage assessment model has passed the training; otherwise, the large-scale damage assessment model continues to be trained. The identification stability evaluation value is calculated using the following formula:

[0015] ;

[0016] In the formula, E i is the identification stability evaluation value of the i-th damage sample group; m is the number of exposure-enhanced damage samples obtained after exposure perturbation enhancement of the original damage sample of a historical accident vehicle; The number of pixels whose damage areas overlap with the j-th exposure-disturbance-enhanced damage sample of the original damage sample of the historical accident vehicle are successfully identified in image recognition. L represents the total number of pixels successfully identified in image recognition of the original damage samples of historical accident vehicles and the total number of pixels successfully identified in image recognition of the damage areas of the exposure-perturbation-enhanced damage samples after deduplication; iL represents the length of consecutive pixels corresponding to the crack skeleton in the original damage sample of historical accident vehicles. ij B represents the length of consecutive pixels corresponding to the crack skeleton in the j-th exposure perturbation enhanced damage sample; i B represents the fracture confidence level in the original damage samples of historical accident vehicles. ij B represents the fracture confidence of the j-th exposure perturbation-enhanced damage sample. i and B ij These are the probability values ​​of the damaged area output by the large-scale damage assessment model, indicating that the damage area belongs to a real fracture.

[0017] In the above-mentioned large model-based damage assessment method, in step S200, the damage assessment large model assesses each damaged area before making an overall damage assessment of the accident vehicle.

[0018] In the aforementioned large-scale model-based damage assessment method, in step S200, the damage assessment large-scale model evaluates each damaged area and characterizes the degree of damage to the damaged area through a comprehensive damage assessment value, which is calculated using the following formula:

[0019] ;

[0020] In the formula, D k P represents the overall damage assessment value for the k-th damaged area; k This represents the number of damaged pixels corresponding to the k-th damaged region. L represents the number of pixels on the vehicle component where the k-th loss region is located; k The length of the continuous pixels corresponding to the crack skeleton in the damage region at the kth location; B represents the scale pixel value of the vehicle component where the k-th damaged area is located; k C represents the fracture confidence level of the damage region at the k-th location; k This represents the preliminary repairability level quantification value corresponding to the k-th damaged area.

[0021] In the above-mentioned loss assessment method based on a large model, in step S200, ...

[0022] In the above-mentioned large model-based damage assessment method, in step S100, when the damage assessment large model assesses the damage based on the identification results, the damage assessment large model outputs repair suggestions and repair costs based on the damage identification results.

[0023] A system for loss assessment using the large-model-based loss assessment method described in claim 1, comprising:

[0024] The data acquisition unit is used to acquire image data and / or video data of the accident vehicle;

[0025] The damage assessment unit is used to assess the damage to the accident vehicle based on the data collected by the data acquisition unit; the damage assessment unit contains a large damage assessment model.

[0026] The storage unit is used to store the data collected by the data acquisition unit and the output results of the damage assessment unit;

[0027] The data acquisition unit is communicatively connected to both the damage assessment unit and the storage unit, and the damage assessment unit is communicatively connected to the storage unit.

[0028] The system also includes a central control unit, which manages and controls the data acquisition unit, damage assessment unit, and storage unit.

[0029] The system also includes a visualization unit that communicates with the central control unit.

[0030] The technical solution of the present invention achieves the following beneficial technical effects:

[0031] 1. The advantages of this application compared to existing technologies are: training a large-scale damage assessment model based on exposure stability evaluation values ​​enables the model to learn stable characteristics of real vehicle damage under different lighting, reflection, shadow, blur, and compressed noise conditions. This reduces the interference of paint reflection, shadow edges, stains, or exposure anomalies on damage identification results, and improves the accuracy and stability of crack area, fracture edge, and damage area identification.

[0032] 2. The advantages of this application compared to existing technologies are: it can further convert image recognition results into repairability levels, recommended repair methods, and estimated price ranges, reducing judgment bias caused by differences in human damage assessment experience, and improving the consistency and reliability of vehicle damage assessment results. Attached Figure Description

[0033] Figure 1 This is a schematic diagram illustrating the working principle of the large-model-based damage assessment system in this invention.

[0034] Figure 2 This is a flowchart of the loss assessment process based on a large model in this invention. Detailed Implementation

[0035] The present invention will be further explained below with reference to examples.

[0036] like Figure 1 As shown, the damage assessment system based on a large model includes a data acquisition unit, a damage assessment unit, a storage unit, a central control unit, and a visualization unit. The central control unit is communicatively connected to the data acquisition unit, the damage assessment unit, the storage unit, and the visualization unit, respectively. The data acquisition unit is communicatively connected to the damage assessment unit and the storage unit, respectively. The damage assessment unit is communicatively connected to the storage unit.

[0037] When assessing damage to accident vehicles, in addition to manual on-site assessment, computer systems are also used. Furthermore, during manual on-site assessment, assessors also take photos and videos of the accident vehicles as evidence. In this invention, the large-scale model-based damage assessment system is a computer system that assesses damage to accident vehicles using images and / or videos.

[0038] In this invention, the data acquisition unit is a shooting device, including a digital camera, a scanning device, a mobile phone with photo and video recording functions, and other electronic devices with photo and video recording functions, used to acquire images (pictures) and video images of the accident vehicle, forming image data and video data.

[0039] As the core of the large-scale model-based damage assessment system, the damage assessment unit is used to assess the damage to accident vehicles based on data collected by the data acquisition unit. The damage assessment unit contains a built-in large-scale damage assessment model, which uses image and video data to identify and assess the damage and repairability of the vehicles. Specifically, the large-scale damage assessment model is a pre-trained model, specifically a large-scale model for accident vehicle damage assessment obtained through deep learning using a multimodal large-scale model.

[0040] Specifically, the large-scale damage assessment model in this invention is trained through the following steps:

[0041] S201) Collect historical accident vehicle image data and historical accident vehicle video data;

[0042] S202) The video image frames of the historical accident vehicle video data collected in step S201 are extracted by combining fixed time interval frame sampling and motion change detection, and the extracted video image frames are filtered to obtain a set of effective video image frames of historical accident vehicles.

[0043] S203) The historical accident vehicle image data collected in step S201 and the effective video image frame set of historical accident vehicles obtained in step S202 are cleaned, deduplicated, and damaged to obtain a set of original damage samples of historical accident vehicles with vehicle damage labels. Then, exposure perturbation enhancement processing is performed on the original damage samples in the set of original damage samples of historical accident vehicles to obtain a dataset of damage samples of historical accident vehicles. The vehicle damage label is one or more of the following: damage type label, damage area label, crack area label, fracture area label, and repairability level label.

[0044] (S204) Divide the historical accident vehicle damage sample dataset obtained in step S203 into a training set, a validation set, and a test set in a ratio of 7:2:1, and use the training set, validation set, and test set to train the damage assessment model.

[0045] In step S203, when marking damage, it is necessary to mark the damage type, repairability level, repair process and repair cost components, and construct a repair rule table that maps the damage type to the repairability level, repair process and repair cost components.

[0046] Given that the quality of historical accident vehicle images and videos is affected by factors such as lighting conditions and shooting angles when collecting historical accident vehicle image data and video data, preprocessing is required before using historical accident vehicle image data and video data for training the large-scale damage assessment model.

[0047] Specifically, historical accident vehicle video data is processed using a combination of fixed-time frame extraction and motion change detection to obtain a dataset of historical accident vehicle video image frames. Then, the sharpness value of each historical accident vehicle video image frame is calculated based on the Laplacian sharpness evaluation algorithm. Combined with damage area visibility analysis and inter-frame structural similarity calculation results, multiple historical accident vehicle video image frames are filtered to remove duplicate frames, blurred frames, and frames where damage areas are not visible, resulting in a valid set of historical accident vehicle video image frames. Inter-frame structural similarity is used to characterize the degree of content repetition between adjacent video image frames to reduce redundant accident image samples; it is calculated using a structural similarity algorithm.

[0048] After obtaining the effective video image frame set of historical accident vehicles, image cleaning processing is performed on the historical accident vehicle image data and the effective video image frame set. Vehicle component regions are extracted using vehicle target detection. Based on the proportion of occluded areas, brightness distribution, and image noise level, severely occluded vehicle component images, abnormally exposed images, and invalid background images are removed to obtain the original damage sample set of historical accident vehicles. Then, a combination of target detection annotation and semantic segmentation annotation is used to annotate the vehicle damage areas in the original damage samples of historical accident vehicles to obtain corresponding vehicle damage labels. The vehicle damage labels include damaged component labels, damage type labels, crack area labels, fracture edge labels, and repairability level labels. In this embodiment, crack area features and fracture edge features are obtained through edge contour extraction algorithms, and the repairability level label is obtained by mapping damage area statistics with a maintenance rule table.

[0049] Due to insufficient data collection or inadequate data categories, training a large model using only existing data would result in insufficient training, leading to inadequate recognition accuracy and failure to meet usage requirements. Therefore, this invention employs exposure perturbation enhancement to augment the labeled original damage samples of historical accident vehicles, thereby obtaining more damage samples while maintaining the original vehicle damage labels. The exposure perturbation enhancement method includes one or more of the following: overexposure enhancement, underexposure enhancement, local shadow overlay, blur perturbation, and compressed noise perturbation.

[0050] After exposure perturbation enhancement, the original damage samples of historical accident vehicles and the exposure perturbation enhanced damage samples are merged into a historical accident vehicle damage sample dataset. Then, the historical accident vehicle damage sample dataset is divided into training set, validation set and test set according to a certain ratio.

[0051] Since the damage area output by the large-scale damage assessment model is usually represented as a pixel-level mask in the image, this invention uses the number of pixels corresponding to the damage area to characterize the area size. For the original damage sample and the exposure disturbance enhanced damage sample of historical accident vehicles, after obtaining their damage area masks, the number of pixels that are identified as damaged in both is counted as the number of overlapping pixels; the number of pixels identified as damaged in both is counted and duplicates are removed to obtain the total number of damaged pixels.

[0052] Since the actual damage area and crack fracture structure of the same vehicle should not substantially change after exposure enhancement, exposure reduction, local shadowing, blurring disturbance, and compression noise disturbance, an exposure stability evaluation value is constructed. When testing the trained damage assessment model, the original damage samples of historical accident vehicles and the exposure-enhanced damage samples obtained after different exposure disturbance enhancement processes are divided into a damage sample group. The trained damage assessment model is used to identify this damage sample group and calculate the identification stability evaluation value. When the identification stability evaluation value is greater than or equal to the preset evaluation threshold, it indicates that the damage assessment model training has reached the target; otherwise, the damage assessment model training continues. The identification stability evaluation value is calculated by the following formula:

[0053] ;

[0054] In the formula, E i is the identification stability evaluation value of the i-th damage sample group; m is the number of exposure-enhanced damage samples obtained after exposure perturbation enhancement of the original damage sample of a historical accident vehicle; The number of pixels whose damage areas overlap with the j-th exposure-disturbance-enhanced damage sample of the original damage sample of the historical accident vehicle are successfully identified in image recognition. L represents the total number of pixels successfully identified in image recognition of the original damage samples of historical accident vehicles and the total number of pixels successfully identified in image recognition of the damage areas of the exposure-perturbation-enhanced damage samples after deduplication; i L represents the length of consecutive pixels corresponding to the crack skeleton in the original damage sample of historical accident vehicles. ij B represents the length of consecutive pixels corresponding to the crack skeleton in the j-th exposure perturbation enhanced damage sample; i B represents the fracture confidence level in the original damage samples of historical accident vehicles. ij B represents the fracture confidence of the j-th exposure perturbation-enhanced damage sample. i and B ij These are the probability values ​​of the damaged area output by the large-scale damage assessment model, indicating that the damage area belongs to a real fracture.

[0055] The greater the overlap between the original damage samples and the exposure-enhanced damage samples of historical accident vehicles, the larger the ratio indicates whether the large damage assessment model still identifies the same damage location after exposure changes. This indicates the consistency of crack length continuity. The closer the crack length of the original damage sample from a historical accident vehicle is to the crack length of the exposure-enhanced damage sample, then... The smaller the value, the closer the crack lengths are, and the closer the ratio is to 1, the more stable the large loss assessment model is. If the difference between the two is large, the ratio will decrease significantly, indicating that the large loss assessment model is less stable. This indicates the consistency of fracture confidence. If the model's judgment of fracture is similar in the original and perturbation plots, this term is larger; if the fracture confidence changes significantly, this term is smaller. or (L) i ,L ij ) max When the value is 0, a preset minimum constant is used for substitution, or the corresponding consistency term is set to 0 to avoid the denominator being 0.

[0056] This means that for the same original damage sample of a historical accident vehicle, it compares it one by one with the corresponding multiple exposure disturbance enhanced damage samples and the original damage sample of the historical accident vehicle, and sums up the stability results obtained from each comparison to comprehensively evaluate the overall identification stability of the damage assessment model for the same damage under different disturbance conditions. 1 / m represents the average of the stability results of all exposure disturbance enhanced damage samples, so that even if different samples correspond to different numbers of exposure disturbance enhanced damage samples, comparable exposure stability evaluation values ​​can be obtained.

[0057] The higher the percentage of overlapping pixels in the damage area between the original damage sample and the exposure-enhanced damage sample from historical accidents, the smaller the difference in crack length and the smaller the difference in fracture confidence, the higher the exposure stability evaluation value, indicating that the damage assessment model's identification result for this damage sample is more stable. Conversely, when the exposure stability evaluation value is low, it indicates that the model is easily affected by paint reflection, shadow occlusion, or exposure abnormalities, and the training constraint weight of such samples needs to be increased during the training of the damage assessment model.

[0058] During the training of the large-scale damage assessment model, the exposure stability evaluation value is used as a training constraint. When the exposure stability evaluation value is lower than 0.4, it indicates that the output result of the large-scale damage assessment model fluctuates greatly under the exposure perturbation conditions corresponding to the original damage sample of the historical accident vehicle. It is easily affected by paint reflection, shadow occlusion, blur perturbation or exposure abnormality. Therefore, the loss weight of the original damage sample of the historical accident vehicle in the next round of training is increased so that the model can further learn the real damage characteristics of this type. When the exposure stability evaluation value is greater than or equal to 0.4, it indicates that the large-scale damage assessment model has good recognition stability of the damage area, crack continuity structure and fracture characteristics in the original damage sample of the historical accident vehicle.

[0059] The pre-trained damage assessment model can stably output vehicle damage area, damage type, crack continuity structure, fracture confidence, number of pixels in the damage area, and preliminary repairability level under different exposure conditions, shooting quality, and damage morphology. By constraining the model tuning process with exposure stability evaluation values, the impact of paint reflection, shadow occlusion, exposure anomalies, blur perturbations, and compression noise on the recognition results can be reduced, providing a model foundation for subsequent damage identification and assessment in vehicle images or videos.

[0060] The trained damage assessment model is then applied to the model-based damage assessment system to assess the damage to accident vehicles. The steps are as follows: Figure 2 As shown, it specifically includes:

[0061] S100) Data Acquisition: Acquire image and video data of the accident vehicle;

[0062] S200) The large-scale damage assessment model is used to identify the image and video data collected in step S100 and to assess the damage based on the identification results. The identification results include the vehicle component area, the damage area mask, the damage type, the number of pixels in the damage area, the continuous pixel length of the crack, the fracture confidence level, and the preliminary repairability level.

[0063] When using the large damage assessment model to identify the image data and video data collected in step S100 and to assess the damage based on the identification results, if unrecognizable image information appears in the image data or unrecognizable video image frames appear in the video data, the corresponding image information or video image frames are output.

[0064] When an accident vehicle image includes multiple images, or an accident vehicle video includes multiple image frames, the damaged areas in different images or different image frames are matched based on the vehicle component area, the center position of the damaged area, the degree of pixel overlap of the damaged area, and the direction of crack extension. Damage identification results belonging to the same vehicle component and similar locations are merged into the same damage target. When an accident vehicle image includes only a single accident image, each damaged area identified in that accident image is considered as a damage target.

[0065] When assessing each damaged area, the large-scale damage assessment model uses a comprehensive damage assessment value to characterize the degree of damage in that area. This comprehensive damage assessment value is calculated using the following formula:

[0066] ;

[0067] In the formula, D k P represents the overall damage assessment value for the k-th damaged area; k This represents the number of damaged pixels corresponding to the k-th damaged region. L represents the number of pixels on the vehicle component where the k-th loss region is located; k The length of the continuous pixels corresponding to the crack skeleton in the damage region at the kth location; B represents the scale pixel value of the vehicle component where the k-th damaged area is located; k C represents the fracture confidence level of the damage region at the k-th location; k This represents the preliminary repairability level quantification value corresponding to the k-th damaged area.

[0068] in, This is used to represent the proportion of damage area at location k relative to the damage range of its corresponding vehicle component. Since the shooting distance, resolution, and vehicle component size may vary between images, using only the number of damaged pixels for judgment can easily lead to magnified damage in close-up images or understated damage in distant images. Therefore, using the ratio of the number of pixels in the damaged area to the number of pixels in the corresponding vehicle component converts the absolute number of pixels into a relative damage range, making the damage range judgment more stable.

[0069] This represents the proportion of the continuous crack length in the k-th damage region relative to the scale of the vehicle component. A larger continuous crack pixel length indicates a higher degree of crack extension; a larger proportion relative to the scale of the vehicle component suggests that the damage is more likely to develop from a simple scratch or localized crack into a more severe crack, fracture, or breakage. Normalizing using vehicle component scale pixel values ​​can reduce the impact of shooting distance, image size, and vehicle component size on crack length determination.

[0070] B k This is used to represent the probability that the damage area at point k is a true fracture. The fracture confidence is output by the fracture identification branch in the damage assessment model. The higher the fracture confidence, the more likely the damage area is to have component fracture, fracture separation, edge cracking, or damage to the structural connection. The corresponding repair methods tend to be component replacement, disassembly and inspection, or inspection of the structural connection area.

[0071] C k This is used to represent the preliminary repairability level of the k-th damaged area. Specifically, different levels such as minor repair, general repair, sheet metal repair, component replacement, or disassembly and inspection can be converted into corresponding quantitative values. The higher the preliminary repairability level, the greater the repair difficulty and cost of the damaged area, and the higher the overall damage assessment value.

[0072] The comprehensive damage assessment formula can be used to comprehensively evaluate the vehicle damage area from four aspects: damage range, crack continuity, fracture probability, and preliminary repairability level. A larger proportion of the damaged area, a larger proportion of the crack continuity length, a higher fracture confidence level, and a higher preliminary repairability level results in a higher comprehensive damage assessment value, indicating that the damage area is more likely to be severely damaged. Conversely, a smaller proportion of the damaged area, a shorter crack continuity length, a lower fracture confidence level, and a lower preliminary repairability level results in a lower comprehensive damage assessment value, indicating that the damage area is more likely to be slightly damaged.

[0073] Furthermore, a corresponding vehicle damage assessment result is generated based on the comprehensive damage assessment value. When the comprehensive damage assessment value is less than the first preset assessment threshold, the k-th damaged area is determined to be slightly damaged, and the corresponding repair methods include polishing, partial painting, or minor repair. When the comprehensive damage assessment value is greater than or equal to the first preset assessment threshold and less than the second preset assessment threshold, the k-th damaged area is determined to be moderately damaged, and the corresponding repair methods include painting repair, sheet metal repair, or partial repair. When the comprehensive damage assessment value is greater than or equal to the second preset assessment threshold, or the fracture confidence level is greater than the preset fracture threshold, the k-th damaged area is determined to be severely damaged, and the corresponding repair methods include component replacement, disassembly and inspection, or structural connection area inspection.

[0074] Furthermore, based on the vehicle component where the k-th damaged area is located, the damage type, the comprehensive damage assessment value, the repairability level, and the recommended repair method, combined with preset repair time parameters, material cost parameters, and component replacement cost parameters, an estimated price range corresponding to that damaged area is generated. If the same target vehicle has multiple damaged areas, the comprehensive damage assessment value, repair method, and estimated price range corresponding to each damaged area are calculated separately. The assessment results for multiple damaged areas are then summarized to obtain the vehicle damage assessment result for the target vehicle.

[0075] After identification by the large-scale damage assessment model, the vehicle damage assessment result can be obtained for the target vehicle. The vehicle damage assessment result includes at least one of the following: damaged components, damage type, damaged area, number of pixels in the damaged area, continuous pixel length of the crack, fracture confidence level, comprehensive damage assessment value, repairability level, recommended repair method, and valuation range. Therefore, based on the damage identification results in the accident images or videos of the vehicle to be assessed, the degree of vehicle damage can be quantitatively assessed, and corresponding repair suggestions and damage assessment results can be generated, improving the stability and consistency of vehicle damage assessment results.

[0076] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of the claims of this patent application.

Claims

1. A loss assessment method based on a large model, characterized in that, include: S100) Data Acquisition: Acquire image and video data of the accident vehicle; S200) The large-scale damage assessment model is used to identify the image and video data collected in step S100 and to assess the damage based on the identification results; wherein, the large-scale damage assessment model is trained through the following steps: S201) Collect historical accident vehicle image data and historical accident vehicle video data; S202) The video image frames of the historical accident vehicle video data collected in step S201 are extracted by combining fixed time interval frame sampling and motion change detection, and the extracted video image frames are filtered to obtain a set of effective video image frames of historical accident vehicles. S203) The historical accident vehicle image data collected in step S201 and the effective video image frame set of historical accident vehicles obtained in step S202 are cleaned, deduplicated, and damaged to obtain a set of original damage samples of historical accident vehicles with vehicle damage labels. Then, exposure perturbation enhancement processing is performed on the original damage samples in the set of original damage samples of historical accident vehicles to obtain a dataset of damage samples of historical accident vehicles. The vehicle damage label is one or more of the following: damage type label, damage area label, crack area label, fracture area label, and repairability level label. (S204) Divide the historical accident vehicle damage sample dataset obtained in step S203 into a training set, a validation set, and a test set in a ratio of 7:2:1, and use the training set, validation set, and test set to train the damage assessment model.

2. The loss assessment method based on a large model according to claim 1, characterized in that, In step S203, the exposure perturbation enhancement is one or more of the following: overexposure enhancement, underexposure enhancement, local shadow overlay, blur perturbation, and compression perturbation, and the original vehicle damage markings remain unchanged when the exposure perturbation enhancement is performed on the original damage samples of historical accident vehicles.

3. The loss assessment method based on a large model according to claim 1, characterized in that, In step S204, when testing the trained damage assessment model, the original damage samples of historical accident vehicles and the exposure-enhanced damage samples obtained by applying different exposure perturbation enhancement processes to the original damage samples of historical accident vehicles are divided into a damage sample group. The trained damage assessment model is used to identify this damage sample group and calculate the identification stability evaluation value. When the identification stability evaluation value is greater than or equal to the preset evaluation threshold, it indicates that the damage assessment model training has reached the target; otherwise, the damage assessment model training continues. The identification stability evaluation value is calculated using the following formula: ； In the formula, E i is the identification stability evaluation value of the i-th damage sample group; m is the number of exposure-enhanced damage samples obtained after exposure perturbation enhancement of the original damage sample of a historical accident vehicle; The number of pixels whose damage areas overlap with the j-th exposure-disturbance-enhanced damage sample of the original damage sample of the historical accident vehicle are successfully identified in image recognition. L represents the total number of pixels successfully identified in image recognition of the original damage samples of historical accident vehicles and the total number of pixels successfully identified in image recognition of the damage areas of the exposure-perturbation-enhanced damage samples after deduplication; i L represents the length of consecutive pixels corresponding to the crack skeleton in the original damage sample of historical accident vehicles. ij B represents the length of consecutive pixels corresponding to the crack skeleton in the j-th exposure perturbation enhanced damage sample; i B represents the fracture confidence level in the original damage samples of historical accident vehicles. ij B represents the fracture confidence of the j-th exposure perturbation-enhanced damage sample. i and B ij These are the probability values ​​of the damaged area output by the large-scale damage assessment model, indicating that the damage area belongs to a real fracture.

4. The loss assessment method based on a large model according to claim 1, characterized in that, In step S200, when assessing the damage of the accident vehicle, the large damage assessment model evaluates each damaged area before making an overall damage assessment.

5. The loss assessment method based on a large model according to claim 4, characterized in that, In step S200, when the large damage assessment model evaluates each damaged area, it characterizes the degree of damage to the damaged area through a comprehensive damage assessment value, which is calculated using the following formula: ； In the formula, D k P represents the overall damage assessment value for the k-th damaged area; k This represents the number of damaged pixels corresponding to the k-th damaged region. L represents the number of pixels on the vehicle component where the k-th loss region is located; k The length of the continuous pixels corresponding to the crack skeleton in the damage region at the kth location; B represents the scale pixel value of the vehicle component where the k-th damaged area is located; k C represents the fracture confidence level of the damage region at the k-th location; k This represents the preliminary repairability level quantification value corresponding to the k-th damaged area.

6. The loss assessment method based on a large model according to claim 1, characterized in that, In step S200, when the large-scale damage assessment model is used to identify the image data and video data collected in step S100 and to assess the damage based on the identification results, if unrecognizable image information appears in the image data or unrecognizable video image frames appear in the video data, the corresponding image information or video image frames are output.

7. The loss assessment method based on a large model according to claim 1, characterized in that, In step S100, when the damage assessment model assesses the damage based on the identification results, it outputs repair suggestions and repair costs based on the damage identification results.

8. A system for loss assessment using the large-model-based loss assessment method described in claim 1, characterized in that, include: The data acquisition unit is used to acquire image data and / or video data of the accident vehicle; The damage assessment unit is used to assess the damage to the accident vehicle based on the data collected by the data acquisition unit. The damage assessment unit contains a large-scale damage assessment model. The storage unit is used to store the data collected by the data acquisition unit and the output results of the damage assessment unit; The data acquisition unit is communicatively connected to both the damage assessment unit and the storage unit, and the damage assessment unit is communicatively connected to the storage unit.

9. The system according to claim 8, characterized in that, It also includes a central control unit, which is used to manage and control the data acquisition unit, damage assessment unit, and storage unit.

10. The system according to claim 8, characterized in that, It also includes a visualization unit that communicates with the central control unit.

Images