An automobile part recognition method, system, electronic device and storage medium

Abstract

The present application relates to the technical field of image recognition, and specifically discloses a vehicle part recognition method and system, an electronic device and a storage medium. The technical scheme of the present application generates an enhanced image dataset by acquiring multi-angle vehicle images and performing image enhancement processing, and divides the fine-tuning process of a pre-trained visual model into three stages to progressively unfreeze and update the top, middle and bottom layer parameters in sequence. Performance evaluation and threshold determination are introduced during stage conversion, solving the problems of excessive consumption of computing resources and excessively long training time caused by full parameter fine-tuning relying on large-scale datasets in the prior art, and the problems of insufficient recognition accuracy and generalization ability caused by the difficulty of pre-trained models to adapt to multi-angle complex backgrounds. The present application reduces the model training cost, shortens the model training cycle, and improves the vehicle part recognition accuracy and generalization ability.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method, system, electronic device, and storage medium for recognizing automotive parts. Background Technology

[0002] In the property insurance claims process, the accuracy of the system's automatic identification of vehicle parts directly impacts claims efficiency and customer satisfaction. Existing computer vision-based technologies primarily rely on large-scale datasets for full-parameter fine-tuning, which consumes significant computational resources and prolongs training time. While some methods alleviate some of these issues through pre-trained models, these methods often struggle to meet the high-precision recognition requirements under multi-angle and complex backgrounds, resulting in both recognition accuracy and processing efficiency failing to meet business requirements in real-world claims scenarios. Furthermore, existing data collection and processing methods typically lack efficient handling of data diversity and noise, leading to insufficient model generalization ability. These technical shortcomings severely restrict the promotion and application of vehicle part recognition technology in the property insurance claims field.

[0003] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method, system, electronic device, and storage medium for identifying automotive parts.

[0005] Firstly, the present invention provides a method for identifying vehicle parts, the technical solution of which is as follows: Acquire multi-angle car images and perform image enhancement processing on the multi-angle car images to generate an enhanced image dataset; A pre-trained visual model is obtained, and the fine-tuning process of the pre-trained visual model is divided into a first stage, a second stage, and a third stage; wherein, the pre-trained visual model includes low-level parameters, mid-level parameters, and top-level parameters; In the first stage, the bottom-level parameters and the middle-level parameters are frozen, and only the top-level parameters are updated. A first round of parameter updates is performed on the enhanced image dataset. After the first stage, a first performance evaluation is performed on the trained visual model. When the first performance evaluation result reaches a first preset performance threshold, the second stage begins. In the second stage, the middle layer parameters are unfrozen while the bottom layer parameters remain frozen. The middle layer parameters and the top layer parameters are updated, and a second round of parameter updates is performed on the enhanced image dataset. After the second stage, a second performance evaluation is performed on the trained visual model. When the second performance evaluation result reaches a second preset performance threshold, the third stage begins. In the third stage, all parameters of the pre-trained visual model are unfrozen, all parameters are updated, and a third round of parameter updates is performed on the augmented image dataset. The output is a target model fine-tuned through the first, second, and third stages, which is used to identify car parts in the input car image.

[0006] The beneficial effects of the vehicle part identification method of the present invention are as follows: The method of this invention acquires multi-angle car images and performs image enhancement processing to generate an enhanced image dataset. It also divides the fine-tuning process of the pre-trained visual model into three stages, progressively unfreezing and updating the parameters of the top, middle and bottom layers in sequence. Performance evaluation and threshold determination are introduced during the stage transition. This solves the problems in the prior art, such as excessive consumption of computing resources and excessively long training time due to the reliance on large-scale datasets for full parameter fine-tuning, and the difficulty of the pre-trained model to adapt to complex multi-angle backgrounds, resulting in insufficient recognition accuracy and generalization ability. This reduces the model training cost, shortens the model training cycle, and improves the recognition accuracy and generalization ability of car parts.

[0007] Based on the above solution, the vehicle part identification method of the present invention can be further improved as follows.

[0008] In one alternative approach, the step of performing image enhancement processing on the multi-angle vehicle images to generate an enhanced image dataset includes: The multi-angle car images are subjected to geometric transformation enhancement, color and lighting enhancement, and background diversification processing to obtain the enhanced image dataset.

[0009] The beneficial effects of adopting the above optional methods are as follows: by further enhancing geometric transformation, color and lighting, and diversifying background processing, the diversity of training data in terms of angle, lighting and scene can be expanded, enabling the model to learn a richer distribution of visual features and enhancing the stability of car parts recognition under different shooting conditions.

[0010] In one alternative approach, prior to the step of performing image enhancement processing on the multi-angle vehicle images, the method further includes: Vehicle sensor data and historical claims image data are collected, and multimodal feature fusion is performed on the multi-angle vehicle images, the vehicle sensor data, and the historical claims image data to generate fused image data. The fused image data is then used as input data for the image enhancement processing.

[0011] The beneficial effects of adopting the above-mentioned optional methods are as follows: Before image enhancement, vehicle sensor data and historical claims image data are introduced to perform multimodal feature fusion, which enriches the information dimension and semantic level of the input data, provides a more comprehensive data foundation for subsequent enhancement processing, and improves the model's adaptability to complex claims scenarios.

[0012] In one alternative approach, the bottom-level parameters are the front transformer layer parameters of the pre-trained visual model, the middle-level parameters are the middle transformer layer parameters of the pre-trained visual model, and the top-level parameters are the back transformer layer parameters and the classification head parameters of the pre-trained visual model.

[0013] The beneficial effects of adopting the above optional approach are as follows: the bottom layer parameters, middle layer parameters and top layer parameters are further mapped to the front, middle and rear transformer layers and classification head of the pre-trained visual model, respectively, so that the three-stage progressive unfreezing strategy is accurately matched with the model structure, and the update timing and range of parameters at different levels are optimized.

[0014] In one alternative approach, in the first stage, the global contour features of the input image are used as input to the attention mechanism, and the output is a feature representation focused on the global contour. In the second stage, the feature representation focused on the global contour is used as the input to the attention mechanism, and the output is an intermediate representation focused on the local component features; In the third stage, the intermediate representation focusing on local component features is used as the input to the attention mechanism, and the output is the target feature representation focusing on detail damage features.

[0015] The beneficial effects of adopting the above optional method are as follows: further extracting global contour features in the first stage, extracting local component features in the second stage, and extracting detailed damage features in the third stage, so that the feature learning granularity and parameter unfreezing depth advance synchronously, and gradually refine the model's ability to perceive vehicle parts and damage.

[0016] In one alternative approach, after the step of outputting the target model, the method further includes: The confidence level of the recognition result obtained by the target model in recognizing the input car image is evaluated. When the confidence level of the recognition result is lower than a preset threshold, a manual review process is triggered.

[0017] The advantages of adopting the above-mentioned optional methods are as follows: a confidence assessment mechanism is further introduced after the model output to determine the reliability of the identification results. When the confidence level is lower than the preset threshold, the manual review process is automatically triggered to avoid low-quality identification results from directly entering the claims process and to ensure the accuracy of business decisions.

[0018] In one alternative approach, after the step that triggers the manual review process, the following is also included: The manually reviewed image data is added to the enhanced image dataset for subsequent incremental training.

[0019] The beneficial effects of adopting the above optional methods are as follows: the manually reviewed image data is further fed back to the enhanced image dataset for subsequent incremental training, continuously expanding the model's learning samples, forming a positive cycle of data accumulation and model optimization, and improving the model's long-term performance in actual business scenarios.

[0020] Secondly, the present invention provides a vehicle part recognition system, the technical solution of which is as follows: The image processing module is used to acquire multi-angle car images and perform image enhancement processing on the multi-angle car images to generate an enhanced image dataset; The stage division module is used to obtain the pre-trained visual model and divide the fine-tuning process of the pre-trained visual model into a first stage, a second stage, and a third stage; wherein, the pre-trained visual model includes bottom-level parameters, middle-level parameters, and top-level parameters; The first update module is used to freeze the bottom-level parameters and the middle-level parameters in the first stage, update only the top-level parameters, and perform a first round of parameter updates on the enhanced image dataset; after the first stage is completed, a first performance evaluation is performed on the trained visual model, and when the first performance evaluation result reaches a first preset performance threshold, the second stage is entered. The second update module is used to unfreeze the middle layer parameters, keep the bottom layer parameters frozen, update the middle layer parameters and the top layer parameters, and perform a second round of parameter updates on the enhanced image dataset in the second stage; after the second stage is completed, a second performance evaluation is performed on the trained visual model, and when the second performance evaluation result reaches a second preset performance threshold, the third stage is entered. The third update module is used in the third stage to unfreeze all parameters of the pre-trained visual model, update all parameters, and perform a third round of parameter updates on the augmented image dataset. The model generation module is used to output a target model after fine-tuning through the first stage, the second stage, and the third stage. The target model is used to identify car parts in the input car image.

[0021] The beneficial effects of the vehicle part recognition system of the present invention are as follows: The system of this invention acquires multi-angle car images and performs image enhancement processing to generate an enhanced image dataset. It also divides the fine-tuning process of the pre-trained visual model into three stages, progressively unfreezing and updating the parameters of the top, middle and bottom layers in sequence. Performance evaluation and threshold determination are introduced during the stage transition. This solves the problems in the prior art, such as excessive consumption of computing resources and excessively long training time due to the reliance on large-scale datasets for full parameter fine-tuning, and insufficient recognition accuracy and generalization ability of the pre-trained model due to its inability to adapt to complex multi-angle backgrounds. The system reduces the model training cost, shortens the model training cycle, and improves the recognition accuracy and generalization ability of car parts.

[0022] Thirdly, the technical solution of an electronic device according to the present invention is as follows: It includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the vehicle part recognition method of the present invention.

[0023] Fourthly, the technical solution of a computer-readable storage medium provided by the present invention is as follows: The computer-readable storage medium stores instructions that, when read, cause the computer-readable storage medium to perform the steps of the vehicle part identification method of the present invention.

[0024] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0025] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating an embodiment of a vehicle part identification method according to the present invention; Figure 2 This is a schematic diagram of the structure of an embodiment of the vehicle part recognition system of the present invention; Figure 3 This is a schematic diagram of an embodiment of an electronic device according to the present invention. Detailed Implementation

[0026] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0027] Figure 1 This diagram illustrates a flowchart of an embodiment of a vehicle part identification method provided by the present invention. This method can be executed by an electronic device such as a terminal device or a server. The terminal device can be any fixed or mobile terminal, such as a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, in-vehicle device, or wearable device. The server can be a single server or a server cluster consisting of multiple servers. Any electronic device can implement the vehicle part identification method by having its processor call computer-readable instructions stored in its memory. Figure 1 As shown, it includes the following steps: S1. Acquire multi-angle car images and perform image enhancement processing on the multi-angle car images to generate an enhanced image dataset.

[0028] Among them, multi-angle car images refer to images that contain the complete or partial appearance of a car, captured from multiple different shooting angles; for example, on the afternoon of May 1, 2026, in parking lot C of city B, nine images of a silver sports utility vehicle of brand A were captured from the front, rear, 45 degrees to the left front, 45 degrees to the right front, 45 degrees to the left rear, 45 degrees to the right rear, the left side, the right side, and from a top-down view.

[0029] Image enhancement processing refers to the process of performing a series of transformations on the original image to increase data diversity and robustness. For example, taking an original image of a silver SUV from brand A at a 45-degree angle from the left front, randomly rotating it by 15 degrees, scaling it to 0.9 times its original size, horizontally flipping it, adjusting the brightness to 1.2 times, and adjusting the contrast to 0.9 times, and then using an image generation model to composite the car body into the road background. An enhanced image dataset refers to a collection of multiple images obtained after image enhancement processing and their corresponding annotation information. For example, 50 different variant images generated from the original image of a silver SUV from brand A at a 45-degree angle from the left front, through geometric transformation, color adjustment, and background replacement, with each image labeled with the shooting angle, car part (left front headlight, front bumper), and damage type (scratches, dents).

[0030] S2. Obtain the pre-trained visual model, and divide the fine-tuning process of the pre-trained visual model into a first stage, a second stage, and a third stage; wherein, the pre-trained visual model includes bottom-level parameters, middle-level parameters, and top-level parameters.

[0031] Among them, a pre-trained visual model refers to a deep learning model that has been pre-trained on a large-scale general image dataset and has the ability to extract basic visual features; for example, a visual transformer model pre-trained on a general image dataset, which contains a 24-layer transformer encoder.

[0032] In this context, bottom-layer parameters refer to the weights and biases of the transformer layers closest to the input layer in the pre-trained visual model; for example, all trainable weights and biases of the transformer encoders from the first to the ninth layer of a 24-layer visual transformer model. Middle-layer parameters refer to the weights and biases of the transformer layers in the middle of the pre-trained visual model; for example, all trainable weights and biases of the transformer encoders from the tenth to the twentieth layer of a 24-layer visual transformer model. Top-layer parameters refer to the weights and biases of the transformer layers closest to the output layer and the classification head in the pre-trained visual model; for example, the weights and biases of the last two transformer encoder layers and the weights and biases of the fully connected classification head in a 24-layer visual transformer model.

[0033] S3. In the first stage, the bottom-level parameters and the middle-level parameters are frozen, and only the top-level parameters are updated. A first round of parameter updates is performed on the enhanced image dataset. After the first stage is completed, a first performance evaluation is performed on the trained visual model. When the first performance evaluation result reaches a first preset performance threshold, the second stage is entered.

[0034] The first performance evaluation result refers to the quantitative index value obtained by testing the current model on the validation set after the first stage of fine-tuning. For example, if 20% of the multi-angle enhanced image dataset of a silver SUV of brand A is used as the validation set, the overall recognition accuracy of the model on the validation set after the first stage of fine-tuning is 86.5%. The first preset performance threshold refers to the minimum performance index value set in advance to determine whether to proceed from the first stage to the second stage; for example, the accuracy threshold is set to 85%.

[0035] S4. In the second stage, the middle layer parameters are unfrozen while the bottom layer parameters are kept frozen. The middle layer parameters and the top layer parameters are updated, and a second round of parameter updates is performed on the enhanced image dataset. After the second stage is completed, a second performance evaluation is performed on the trained visual model. When the second performance evaluation result reaches a second preset performance threshold, the third stage is entered.

[0036] The second performance evaluation result refers to the quantitative index value obtained by testing the current model on the validation set after the second stage of fine-tuning; for example, the overall recognition accuracy of the model on the same validation set after the second stage of fine-tuning is 92.3%. The second preset performance threshold refers to the minimum performance index value set in advance to determine whether to enter the third stage from the second stage; for example, the accuracy threshold is set to 92%.

[0037] S5. In the third stage, all parameters of the pre-trained visual model are unfrozen, all parameters are updated, and a third round of parameter updates is performed on the augmented image dataset.

[0038] S6. Output the target model after fine-tuning through the first stage, the second stage and the third stage, the target model being used to identify car parts in the input car image.

[0039] The target model refers to the final visual model that can be directly used for vehicle part recognition, obtained after incremental fine-tuning in the first, second, and third stages; for example, the visual transformer model after three stages of fine-tuning on a multi-angle enhanced image dataset of a silver sports utility vehicle of brand A.

[0040] Here, "vehicle image" refers to any captured image of a vehicle's exterior that needs to be identified; for example, a photo of a silver Brand A SUV with damage at a 45-degree angle to the left front, uploaded by a user through a claims application, taken on the evening of May 1, 2026, in Parking Lot C of City B. "Vehicle part" refers to a distinguishable component or area on the exterior surface of a vehicle; for example, the left front headlight, front bumper, hood, left front fender, left front door, left rear door, left rear fender, rear bumper, right front headlight, and right front fender.

[0041] The technical solution of this embodiment acquires multi-angle car images and performs image enhancement processing to generate an enhanced image dataset. It also divides the fine-tuning process of the pre-trained visual model into three stages, progressively unfreezing and updating the parameters of the top, middle and bottom layers in sequence. Performance evaluation and threshold determination are introduced during the stage transition. This solves the problems in the prior art, such as excessive consumption of computing resources and excessively long training time due to the reliance on large-scale datasets for full parameter fine-tuning, and the difficulty of the pre-trained model to adapt to complex multi-angle backgrounds, resulting in insufficient recognition accuracy and generalization ability. This reduces the model training cost, shortens the model training cycle, and improves the recognition accuracy and generalization ability of car parts.

[0042] In one alternative approach, the step of performing image enhancement processing on the multi-angle vehicle images to generate an enhanced image dataset includes: The multi-angle car images are subjected to geometric transformation enhancement, color and lighting enhancement, and background diversification processing to obtain the enhanced image dataset.

[0043] Geometric transformation enhancement refers to data augmentation methods that perform spatial transformations such as rotation, scaling, translation, and flipping on images to change their geometric shape. For example, a new training sample can be generated by randomly rotating the original image of a silver SUV (Brand A) at the front left 45 degrees by 15 degrees, scaling it to 0.8 times, translating it to the left by 10% of its width, and flipping it horizontally. Color and lighting enhancement refers to data augmentation methods that adjust the color attributes of an image, such as brightness, contrast, and saturation, to simulate different lighting conditions. For example, the brightness of the original image of a silver SUV (Brand A) at the front left 45 degrees can be adjusted to 1.3 times, the contrast to 1.2 times, and the saturation to 0.8 times to simulate the lighting effect at dusk. Background diversification refers to a data augmentation method that uses an image generation model to composite the vehicle subject into various different scene backgrounds to increase background complexity and diversity. For example, an image generation model can be used to generate five backgrounds: street, parking lot, road, snow, and rain. The subject of a silver SUV of brand A can then be composited into each of these backgrounds to generate five different training images.

[0044] Among the above-mentioned optional methods, further enhancements such as geometric transformation, color and lighting enhancement, and background diversification are used to expand the diversity of training data in terms of angle, lighting, and scene, enabling the model to learn a richer distribution of visual features and enhancing the stability of car parts recognition under different shooting conditions.

[0045] In one alternative approach, prior to the step of performing image enhancement processing on the multi-angle vehicle images, the method further includes: Vehicle sensor data and historical claims image data are collected, and multimodal feature fusion is performed on the multi-angle vehicle images, the vehicle sensor data, and the historical claims image data to generate fused image data. The fused image data is then used as input data for the image enhancement processing.

[0046] Vehicle sensor data refers to auxiliary information collected from sensors such as the vehicle's inertial measurement unit, global positioning system, and onboard cameras; for example, the vehicle speed, acceleration, heading angle, latitude and longitude coordinates, and dashcam video frames simultaneously recorded when a silver SUV of Brand A is photographed at a 45-degree angle to the left front. Historical claims image data refers to historical vehicle damage images and their annotation information accumulated in property insurance claims, which have been manually annotated or verified through case verification; for example, 100,000 claims images collected in the past year that have been annotated with vehicle parts and damage types, each image containing three annotation fields: shooting angle, damage location, and damage level.

[0047] The fused image data refers to the joint data containing multimodal information generated by feature-level fusion of multi-angle vehicle images, vehicle sensor data, and historical claims image data. For example, attention-weighted fusion is performed on the original image of the left front 45-degree angle of a silver A-brand SUV, the GPS coordinates and inertial measurement unit acceleration data recorded synchronously by the vehicle, and the annotation information of similar angles and similar models in historical claims data to generate a fused image with enhanced position and semantic features.

[0048] Among the above-mentioned optional methods, vehicle sensor data and historical claims image data are further introduced before image enhancement to perform multimodal feature fusion, enriching the information dimension and semantic level of the input data, providing a more comprehensive data foundation for subsequent enhancement processing, and improving the model's adaptability to complex claims scenarios.

[0049] In one alternative approach, the bottom-level parameters are the front transformer layer parameters of the pre-trained visual model, the middle-level parameters are the middle transformer layer parameters of the pre-trained visual model, and the top-level parameters are the back transformer layer parameters and the classification head parameters of the pre-trained visual model.

[0050] The parameters of the front transformer layers refer to all trainable parameters contained in the first few consecutive transformer layers in the pre-trained visual model; for example, the attention weights, feedforward network weights, layer normalization parameters and biases in the encoders of the first to ninth layers of a 24-layer visual transformer model.

[0051] The parameters of the middle transformer layer refer to all trainable parameters contained in several consecutive transformer layers located in the middle of the pre-trained visual model; for example, the attention weights, feedforward network weights, layer normalization parameters and biases in the transformer encoders of the 10th to 20th layers of a 24-layer visual transformer model.

[0052] The parameters of the later transformer layers refer to all trainable parameters contained in the later successive transformer layers of the pre-trained visual model; for example, the attention weights, feedforward network weights, layer normalization parameters, and biases in the encoder layers 21 to 24 of a 24-layer visual transformer model. The parameters of the classification head refer to all trainable weights and biases contained in the fully connected classification layer at the end of the pre-trained visual model; for example, the weight matrix and bias vector in the classification head output by the 24-layer visual transformer model used to map 2048-dimensional features to 9 shooting angle categories and 20 vehicle part categories.

[0053] In the above-mentioned optional methods, the bottom-level parameters, middle-level parameters, and top-level parameters are further mapped to the front, middle, and rear transformer layers and the classification head of the pre-trained visual model, respectively, so that the three-stage progressive unfreezing strategy is accurately matched with the model structure, and the update timing and range of parameters at different levels are optimized.

[0054] In one alternative approach, in the first stage, the global contour features of the input image are used as input to the attention mechanism, and the output is a feature representation focused on the global contour.

[0055] Global contour features refer to features describing the overall shape of a car, its circumscribed rectangle, aspect ratio, and edge contours, among other macroscopic geometric attributes. For example, extracting the minimum circumscribed rectangle, aspect ratio, and convex hull edge point set of the car's contour from a 45-degree left front image of a silver A-brand SUV. Attention mechanisms refer to weighted aggregation vectors in the transformer model by calculating the similarity between the query vector and the key vector, thereby dynamically focusing the computational modules on important regions of the input features. For example, in the transformer layer of a visual transformer model, the attention mechanism calculates the similarity between each image patch and all other image patches, generating a weighted feature map.

[0056] Among them, the feature representation focusing on the global contour refers to the feature vector or feature map that highlights the overall shape information of the car after being processed by the attention mechanism; for example, after the first stage of attention calculation, the feature map of a silver sports utility vehicle from the left front 45 degrees has high activation values ​​concentrated on the outer edge of the vehicle body and the roof outline, while the activation values ​​of the internal texture details are low.

[0057] In the second stage, the feature representation focused on the global contour is used as the input to the attention mechanism, and the output is an intermediate representation focused on the features of local components.

[0058] Among them, the intermediate representation focusing on local component features refers to the feature expression obtained by further strengthening the information of various component areas of the car through an attention mechanism on the basis of global contour features. For example, after the second stage of attention calculation, the feature representation of the left front 45-degree image of a silver sports utility vehicle of brand A, which focuses on the global contour, shows that the areas with higher activation values ​​in the feature map are concentrated in the component areas such as the left front headlight, front bumper, left front fender, and engine hood.

[0059] In the third stage, the intermediate representation focusing on local component features is used as the input to the attention mechanism, and the output is the target feature representation focusing on detail damage features.

[0060] Among them, the target feature representation focusing on detailed damage features refers to the final feature expression obtained by further enhancing the micro-texture information of the damaged area through an attention mechanism based on the local component features. For example, after the third stage of attention calculation, the intermediate representation of the left front 45-degree image of a silver sports utility vehicle of brand A, which focuses on local component features, shows that the area with the highest activation value in the feature map is concentrated on a scratch on the front bumper and a dent at the edge of the left front headlight.

[0061] In the above-mentioned optional methods, global contour features are further extracted in the first stage, local component features are extracted in the second stage, and detailed damage features are extracted in the third stage. This allows the feature learning granularity and parameter unfreezing depth to progress synchronously, gradually refining the model's ability to perceive vehicle parts and damage.

[0062] In one alternative approach, after the step of outputting the target model, the method further includes: The confidence level of the recognition result obtained by the target model in recognizing the input car image is evaluated. When the confidence level of the recognition result is lower than a preset threshold, a manual review process is triggered.

[0063] The recognition result refers to the classification labels and corresponding probabilities output by the target model after reasoning about the input car image, including the shooting angle, car part, and damage type. For example, the target model outputs the following for a photo of a silver SUV with damage at a 45-degree angle to the left front: shooting angle: 45 degrees to the left front (probability 0.92), car part: front bumper (probability 0.87), damage type: scratch (probability 0.81) for a photo of a silver SUV with damage at a 45-degree angle to the left front: Confidence score refers to the highest category probability value output by the target model for a given recognition result, used to measure the reliability of the recognition result. For example, if the target model outputs a front bumper recognition probability of 0.87 for a photo of a silver SUV with damage at a 45-degree angle to the left front, then the confidence score of this recognition result is 0.87.

[0064] The preset threshold refers to a pre-set confidence level threshold used to determine whether the recognition result is reliable and requires manual review; for example, the set confidence level threshold is 0.85. The manual review process refers to the operation where, when the confidence level of the model recognition result is lower than the preset threshold, the image and model output are handed over to a human annotator for secondary judgment and correction; for example, for a blurry image of a silver brand A SUV taken at night, with a model output confidence level of 0.72, the system automatically sends the image and model output to the manual annotation platform, where annotators manually annotate the vehicle parts and damage types on the screen.

[0065] Among the above-mentioned optional methods, a confidence assessment mechanism is further introduced after the model output to determine the reliability of the identification results. When the confidence level is lower than the preset threshold, the manual review process is automatically triggered to avoid low-quality identification results from directly entering the claims process and to ensure the accuracy of business decisions.

[0066] In one alternative approach, after the step that triggers the manual review process, the following is also included: The manually reviewed image data is added to the enhanced image dataset for subsequent incremental training.

[0067] The manually reviewed image data refers to images that have been re-annotated or corrected by human annotators, along with their corresponding correct annotation information. For example, a human annotator might re-annotate a blurry nighttime image of a silver Brand A SUV as having a dent in the left rear door, and save this annotation along with the image as a reviewed data entry. Incremental training refers to adding newly acquired labeled data to the existing training dataset and performing additional rounds of fine-tuning on the existing model to continuously optimize its performance. For example, adding 1000 manually reviewed images of silver Brand A SUVs and their correct annotations to the augmented image dataset, and then retraining the target model for one more round using the same three-stage fine-tuning method, improves the model's accuracy in recognizing low-light nighttime scenes from 72% to 86%.

[0068] In the above-mentioned optional methods, the manually reviewed image data is further fed back to the enhanced image dataset for subsequent incremental training, continuously expanding the model's learning samples, forming a positive cycle of data accumulation and model optimization, and improving the model's long-term performance in real business scenarios.

[0069] To achieve efficient and accurate vehicle part identification, this embodiment proposes an incremental fine-tuning method based on a pre-trained visual model. This method constructs a comprehensive dataset containing nine angle categories—front, rear, left front 45 degrees, right front 45 degrees, left rear 45 degrees, right rear 45 degrees, left side, right side, and other angles—through multi-angle data acquisition and synthesis. In data processing, a self-supervised learning method is employed, combined with image enhancement techniques to generate diverse samples, thereby improving the model's generalization ability.

[0070] During training, an incremental fine-tuning strategy is introduced, dividing the entire fine-tuning process into multiple stages. Each stage updates only a subset of parameters, such as the weights or biases corresponding to a specific layer. After each stage, the model performance is evaluated to determine whether to proceed to the next stage, thereby reducing computational resource requirements and shortening training time. Simultaneously, transfer learning is incorporated, utilizing the knowledge already acquired by the pre-trained model to further improve training efficiency and model accuracy.

[0071] In practice, firstly, multi-angle car images are collected and synthesized to construct a comprehensive dataset containing nine angle categories. Then, a self-supervised learning method is applied, combined with image enhancement techniques, to generate diverse samples and improve dataset diversity. The image enhancement strategies employed include geometric transformation enhancement, color and lighting enhancement, and background complication processing. Geometric transformation enhancement involves applying random rotation (±15 degrees), scaling (0.8 to 1.2 times), translation (±10%), and horizontal flipping to each original image, generating 3 to 5 variant samples for each image. Color and lighting enhancement involves adjusting the image's brightness (0.7 to 1.3 times), contrast (0.8 to 1.2 times), and saturation (0.8 to 1.2 times) to simulate different lighting conditions such as morning, noon, dusk, and night. Background complication processing involves generating diverse background scenes such as city streets, parking lots, and rural roads using an image generation model (e.g., Stable Diffusion), and then synthesizing the vehicle subjects into these complex backgrounds.

[0072] The fine-tuning process is divided into three progressive stages, with each stage updating parameters differently for different layers of the pre-trained visual model. The first stage is the basic feature adaptation stage: freezing the low-level and mid-level parameters of the pre-trained visual model, and only fine-tuning the last 2 to 3 layers of the transformer encoder and the classifier head, with the learning rate set to 1e. -4 The training process consists of two epochs, aiming to adapt the model to the basic visual features of car images. The second stage is the mid-layer semantic optimization stage: the intermediate layer parameters (e.g., layers 10 to 20) are unfrozen, while the bottom layer parameters remain frozen, and the learning rate is reduced to 5e. -5 The training process consists of three epochs, focusing on optimizing the model's semantic understanding of automotive parts (e.g., distinguishing different parts and recognizing damage features). The third stage is the full parameter fine-tuning stage: all parameters are fully unfrozen, and all parameters are fine-tuned, with the learning rate further reduced to 1e. -5 The model is trained for 1 to 2 epochs to converge to its optimal performance. After each epoch, the model performance is evaluated, and the decision to proceed to the next epoch is based on the evaluation results.

[0073] In terms of multimodal data acquisition and processing, this embodiment adopts a multimodal fusion strategy, integrating three major data sources: image data, sensor data, and historical labeled data. A high-quality training dataset is generated through a systematic data processing pipeline. Specifically, data acquisition begins with the collection of raw data through three channels: multi-angle image acquisition (covering nine shooting angles: front, rear, left front 45 degrees, right front 45 degrees, left rear 45 degrees, right rear 45 degrees, left side, right side, and others); vehicle sensor data (including inertial measurement unit data, GPS data, and vehicle camera data); and historical claims image data (from labeled claims case image datasets). Initial processing then proceeds: preprocessing operations are performed on the three types of data, namely, denoising, cropping, and normalizing the image data; spatiotemporal alignment processing for the sensor data; and quality filtering and deduplication processing for the labeled data. Next, multimodal feature fusion is performed: the processed three types of data are fused at the feature level, using an attention-weighted mechanism to weight and fuse features from different modalities, generating a unified multimodal feature representation. Data augmentation is then performed: geometric transformations (rotation, scaling, translation, flipping), color and lighting enhancements (brightness, contrast, saturation adjustment), and background diversification (generating and compositing diverse background scenes using an image generation model) are applied to the fused data. Finally, quality verification is performed: image sharpness checks, annotation consistency verification, and class balance checks are conducted on the augmented data to ensure the quality and diversity of the final training dataset.

[0074] The detailed process of the three-stage incremental fine-tuning algorithm is as follows in this embodiment: First, a pre-trained visual model (e.g., Qwen3.5-4B or Qwen3.5-9B) is loaded as the base model, and the pre-trained weights are retained as initialization parameters. In the first stage (basic feature adaptation), the low-level parameters (layers 1 to 9) and mid-level parameters (layers 10 to 20) of the model are frozen, and only the last 2 to 3 layers of the transformer encoder and classifier head are fine-tuned, using a large learning rate of 1e. -4 The model is trained for two epochs to quickly adapt to the basic visual feature distribution of car images. Freezing the low-level parameters helps preserve the general visual features already learned by the pre-trained model, avoiding overfitting on small sample datasets. After the first stage, the model accuracy is evaluated on the validation set, with a first preset performance threshold of 85%. If the accuracy does not reach this threshold, the hyperparameters are adjusted (e.g., increasing the amount of training data, adjusting the learning rate, or increasing the number of training epochs), and the model is retrained in the first stage. If the threshold is reached, the model proceeds to the second stage. In the second stage (mid-layer semantic optimization), the intermediate layer parameters (layers 10 to 20) are unfrozen while the low-level parameters remain frozen, using a smaller learning rate of 5e. -5The training process consists of three epochs, focusing on optimizing the model's semantic understanding of automotive parts (including differentiation of different parts and identification of damage features). After the second phase, the model's accuracy is evaluated again on the validation set, with a second preset performance threshold of 92%. If the accuracy does not reach this threshold, the hyperparameters are adjusted (e.g., unfreezing more layers or further reducing the learning rate), and the training returns to the second phase for retraining. If the threshold is reached, the training proceeds to the third phase. In the third phase (full parameter fine-tuning), all parameters are unfrozen for full parameter fine-tuning using the minimum learning rate 1e. -5 The model is trained for 1 to 2 epochs, and global parameter optimization is performed to bring it to optimal performance. Using a very small learning rate avoids destructive updates to already optimized parameters. Finally, the final trained model weights are exported, quantized and compressed (e.g., INT8 quantization), and deployed to the inference service.

[0075] The adaptive learning rate adjustment strategy is given by the following formula: in The baseline learning rate is 1e -4 , The attenuation factor is 0.5. This is the number of the current stage (values ​​can be 1, 2, or 3).

[0076] The attention enhancement mechanism involves introducing differentiated attention weights at each stage. The first stage focuses on global contour features, the second on local component features, and the third on detailed damage features. By dynamically adjusting the weights of each attention head at different stages through a multi-head attention mechanism, the model focuses on the most relevant feature regions at different training phases.

[0077] In the vehicle part recognition reasoning and claims application process, after the model training is completed, this embodiment performs the following operations in a real property insurance claims scenario. First, image upload: The user uploads photos of vehicle damage via the claims application, or the claims service robot automatically collects the images. Then, image preprocessing: Standardized preprocessing is performed on the uploaded images, including size normalization (unified to the model input size, e.g., 448×448 pixels), color normalization, and automatic cropping of the region of interest. Next, model inference: The preprocessed image is input into the incrementally fine-tuned pre-trained visual model, and the model outputs recognition results, including shooting angle classification (9 categories), vehicle part recognition, and damage type detection. Then, confidence evaluation: A confidence evaluation is performed on the model's output recognition results. If the highest category probability (i.e., confidence) is greater than or equal to 0.85, the recognition result is directly output; if the confidence is lower than 0.85, a manual review process is triggered. Finally, the results are output and damage assessment is performed: For high-confidence identification results, the process directly proceeds to the claims assessment stage, automatically generating a comprehensive assessment report that includes damage level, repair recommendations, and cost estimates; for low-confidence identification results, the process is transferred to human annotators for review, and the human annotation results are simultaneously sent back to the dataset for subsequent incremental training, thus forming a feedback loop for continuous learning.

[0078] In another embodiment of the vehicle part identification method of the present invention, the following steps are specifically included: S10. Obtain multi-angle vehicle images collected during vehicle insurance claim reporting, and obtain the corresponding vehicle identification code. Based on the vehicle identification code, extract the standard component location mapping table of the vehicle from the preset vehicle model database.

[0079] The Vehicle Identification Number (VIN) is a string code used to uniquely identify a vehicle. For example, a silver A-brand SUV photographed on May 1, 2026, has the VIN L6T7890X3G1234567. The vehicle model database is a pre-set collection of data storing structural parameters, component location distribution, and damage characteristics of different vehicle models. For example, a database containing 50 vehicle models from brands A, B, and C, where each model records the coordinate range of 20 components such as the front bumper, left headlight, and hood in a standard coordinate system.

[0080] The standard component location mapping table refers to a mapping table extracted from the vehicle model database based on the vehicle identification number (VIN) to describe the expected location range of each component of a specific vehicle in an image. For example, the standard component location mapping table of a silver A-brand SUV corresponding to the vehicle identification number L6T7890X3G1234567 can be obtained from the vehicle model database. The image coordinate range of the left front headlight is 100 to 180 pixels on the horizontal axis and 120 to 200 pixels on the vertical axis, and the coordinate range of the front bumper is 80 to 240 pixels on the horizontal axis and 160 to 280 pixels on the vertical axis.

[0081] S20. Concatenate the features of the multi-angle car images and historical images with the same vehicle identification number from the historical claims image data to generate fused image data. Perform geometric transformation enhancement, color and lighting enhancement, and background diversification processing on the fused image data to obtain an enhanced image dataset.

[0082] S30. Obtain the pre-trained visual model. Divide the fine-tuning process of the pre-trained visual model into the first stage, the second stage, and the third stage. The pre-trained visual model includes bottom-level parameters, middle-level parameters, and top-level parameters.

[0083] S40. In the first stage, the bottom-level and middle-level parameters are frozen, and only the top-level parameters are updated. The first round of parameter updates is performed on the augmented image dataset. At the same time, the standard component position mapping table of the vehicle is used as a spatial prior constraint so that the top-level parameter update focuses on the region in the image corresponding to the component position.

[0084] S50. After the first stage is completed, the visual model under training is subjected to a first performance evaluation. When the first performance evaluation result reaches the first preset performance threshold, the second stage begins.

[0085] S60. In the second stage, the middle layer parameters are unfrozen while the bottom layer parameters are kept frozen. The middle layer and top layer parameters are updated, and a second round of parameter updates is performed on the enhanced image dataset. At the same time, the standard part position mapping table is converted into an attention bias matrix to guide the middle layer parameters to focus on the relative positional relationships between parts.

[0086] The attention bias matrix refers to the matrix used to adjust the similarity score between the query vector and the key vector during the attention mechanism calculation process, in order to guide attention to focus on a specific spatial region. For example, a 448×448 attention bias matrix is ​​generated based on a standard component location mapping table. The offset is increased by 0.5 in the coordinate region corresponding to the left front headlight and decreased by 0.2 in other regions, so that the component region is given priority during attention calculation.

[0087] S70. After the second stage, a second performance evaluation is performed on the visual model in training. When the second performance evaluation result reaches the second preset performance threshold, the third stage begins.

[0088] S80. In the third stage, all parameters of the pre-trained visual model are unfrozen, all parameters are updated, and a third round of parameter updates is performed on the enhanced image dataset. At the same time, a damage category balance loss function based on vehicle identification code is introduced to adaptively weight and compensate for the differences in damage type distribution among different vehicle models in the training samples. The finely tuned target model is then output for vehicle part identification in auto insurance claims.

[0089] The damage category balancing loss function refers to a loss function that weights and adjusts the classification errors of different damage types according to their sample distribution during model training, thus mitigating the problem of sample class imbalance. For example, for the three damage types of scratches, dents, and cracks, where scratches account for 80%, dents for 15%, and cracks for 5% of the samples in the training set, the damage category balancing loss function is calculated as scratch error multiplied by a weight of 0.4, dent error multiplied by a weight of 1.2, and crack error multiplied by a weight of 3.0. Adaptive weighted compensation refers to the process of dynamically adjusting the weights of each category in the loss function based on the vehicle model characteristics corresponding to the vehicle identification code of the input image and the distribution of that model in the training samples. For example, for the vehicle identification code L6T7890X3G1234567 corresponding to a silver A-brand SUV, this model only accounts for 10% of the samples in the training set, while a certain B-brand model accounts for 60%. Adaptive weighted compensation adjusts the loss weight of the A-brand samples from 1.0 to 2.5, allowing the model to obtain a larger gradient update for the recognition error of the rare model.

[0090] The technical solution of this embodiment obtains the vehicle identification code and extracts the standard component location mapping table. After feature concatenation of multi-angle vehicle images and historical images of the same model, geometric transformation enhancement, color and lighting enhancement, and background diversification processing are performed to generate an enhanced image dataset. In the three-stage incremental fine-tuning process, the standard component location mapping table is used as a spatial prior constraint to make the top-level parameter update focus on the corresponding region of the component. The standard component location mapping table is converted into an attention bias matrix to guide the middle-level parameters to focus on the relative positional relationship of the components. A damage category balance loss function based on the vehicle identification code is introduced to adaptively weight and compensate for the differences in the distribution of damage types of different models. This solves the problems in the prior art that ignore the model's prior information, resulting in low model recognition accuracy for rare models, insufficient cross-model generalization ability, and classification bias caused by imbalanced damage category samples. It realizes refined part recognition based on model adaptive guidance, improves the accuracy of vehicle part recognition and the model's generalization ability for different models under the conditions of multiple models, multiple angles, and multiple damage types in the car insurance claims scenario.

[0091] Figure 2 A schematic diagram of an embodiment of a vehicle part recognition system 200 provided by the present invention is shown. Figure 2 As shown, the vehicle part recognition system 200 includes: Image processing module 201 is used to acquire multi-angle car images and perform image enhancement processing on the multi-angle car images to generate an enhanced image dataset; The stage division module 202 is used to obtain the pre-trained visual model and divide the fine-tuning process of the pre-trained visual model into a first stage, a second stage, and a third stage; wherein the pre-trained visual model includes bottom-level parameters, middle-level parameters, and top-level parameters. The first update module 203 is used to freeze the bottom layer parameters and the middle layer parameters in the first stage, update only the top layer parameters, and perform a first round of parameter updates on the enhanced image dataset; after the first stage is completed, a first performance evaluation is performed on the visual model under training, and when the first performance evaluation result reaches a first preset performance threshold, the second stage is entered. The second update module 204 is used to unfreeze the middle layer parameters, keep the bottom layer parameters frozen, update the middle layer parameters and the top layer parameters, and perform a second round of parameter updates on the enhanced image dataset in the second stage; after the second stage is completed, a second performance evaluation is performed on the visual model under training, and when the second performance evaluation result reaches a second preset performance threshold, the third stage is entered. The third update module 205 is used to unfreeze all parameters of the pre-trained visual model, update all parameters, and perform a third round of parameter updates on the augmented image dataset in the third stage. The model generation module 206 is used to output a target model after fine-tuning through the first stage, the second stage and the third stage, the target model being used to identify car parts in the input car image.

[0092] In an alternative embodiment, the image processing module 201 is specifically used for: The multi-angle car images are subjected to geometric transformation enhancement, color and lighting enhancement, and background diversification processing to obtain the enhanced image dataset.

[0093] In an alternative embodiment, the image processing module 201 is further configured to: Vehicle sensor data and historical claims image data are collected, and multimodal feature fusion is performed on the multi-angle vehicle images, the vehicle sensor data, and the historical claims image data to generate fused image data. The fused image data is then used as input data for the image enhancement processing.

[0094] In one alternative approach, the bottom-level parameters are the front transformer layer parameters of the pre-trained visual model, the middle-level parameters are the middle transformer layer parameters of the pre-trained visual model, and the top-level parameters are the back transformer layer parameters and the classification head parameters of the pre-trained visual model.

[0095] In one alternative approach, in the first stage, the global contour features of the input image are used as input to the attention mechanism, and the output is a feature representation focused on the global contour. In the second stage, the feature representation focused on the global contour is used as the input to the attention mechanism, and the output is an intermediate representation focused on the local component features; In the third stage, the intermediate representation focusing on local component features is used as the input to the attention mechanism, and the output is the target feature representation focusing on detail damage features.

[0096] In one alternative approach, it also includes: The model verification module is used to evaluate the confidence level of the recognition result obtained by the target model in recognizing the input car image. When the confidence level of the recognition result is lower than a preset threshold, the manual verification process is triggered.

[0097] In one alternative approach, it also includes: The incremental training module is used to add manually reviewed image data to the enhanced image dataset for subsequent incremental training.

[0098] It should be noted that the beneficial effects of the vehicle part recognition system 200 provided in the above embodiments are the same as those of the vehicle part recognition method described above, and will not be repeated here. Furthermore, the system provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to the actual situation to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiments, and will not be repeated here.

[0099] The vehicle part recognition system 200 of the present invention may be a computer program (including program code) running on a computer device. For example, the vehicle part recognition system 200 of the present invention is an application software that can be used to execute the corresponding steps in the vehicle part recognition method of the present invention.

[0100] In some embodiments, the vehicle part recognition system 200 of the present invention can be implemented in a combination of hardware and software. As an example, the vehicle part recognition system 200 of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the vehicle part recognition method of the present invention. For example, the processor in the form of a hardware decoding processor can be one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0101] The modules described in the embodiments of this invention can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.

[0102] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-described vehicle part identification methods. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the vehicle part identification method shown in any embodiment of the present invention by calling the computer program.

[0103] In one alternative embodiment, an electronic device is provided, such as Figure 3 As shown, Figure 3 The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0104] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0105] Bus 4002 may include a path for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus 4002 is represented by only one thick line, but this does not mean that there is only one bus or one type of bus.

[0106] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0107] The memory 4003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.

[0108] Among them, electronic devices can also be terminal devices. A terminal device can be any terminal device that can install applications and access web pages through applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0109] It should be noted that, Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0110] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-described methods for identifying vehicle parts.

[0111] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0112] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned vehicle part identification method.

[0113] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0114] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0115] The computer-readable storage medium provided in this invention can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0116] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0117] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0118] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0119] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0120] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for identifying vehicle parts, characterized in that, include: Acquire multi-angle car images and perform image enhancement processing on the multi-angle car images to generate an enhanced image dataset; A pre-trained visual model is obtained, and the fine-tuning process of the pre-trained visual model is divided into a first stage, a second stage, and a third stage; wherein, the pre-trained visual model includes low-level parameters, mid-level parameters, and top-level parameters; In the first stage, the bottom-level parameters and the middle-level parameters are frozen, and only the top-level parameters are updated. A first round of parameter updates is performed on the enhanced image dataset. After the first stage, a first performance evaluation is performed on the trained visual model. When the first performance evaluation result reaches a first preset performance threshold, the second stage begins. In the second stage, the middle layer parameters are unfrozen while the bottom layer parameters remain frozen. The middle layer parameters and the top layer parameters are updated, and a second round of parameter updates is performed on the enhanced image dataset. After the second stage, a second performance evaluation is performed on the trained visual model. When the second performance evaluation result reaches a second preset performance threshold, the third stage begins. In the third stage, all parameters of the pre-trained visual model are unfrozen, all parameters are updated, and a third round of parameter updates is performed on the augmented image dataset. The output is a target model fine-tuned through the first, second, and third stages, which is used to identify car parts in the input car image.

2. The vehicle part identification method according to claim 1, characterized in that, The steps of performing image enhancement processing on the multi-angle car images to generate an enhanced image dataset include: The multi-angle car images are subjected to geometric transformation enhancement, color and lighting enhancement, and background diversification processing to obtain the enhanced image dataset.

3. The vehicle part identification method according to claim 2, characterized in that, Before the step of performing image enhancement processing on the multi-angle vehicle images, the method further includes: Vehicle sensor data and historical claims image data are collected, and multimodal feature fusion is performed on the multi-angle vehicle images, the vehicle sensor data, and the historical claims image data to generate fused image data. The fused image data is then used as input data for the image enhancement processing.

4. The vehicle part identification method according to claim 1, characterized in that, The bottom layer parameters are the parameters of the front transformer layer of the pre-trained visual model, the middle layer parameters are the parameters of the middle transformer layer of the pre-trained visual model, and the top layer parameters are the parameters of the back transformer layer and the classification head parameters of the pre-trained visual model.

5. The vehicle part identification method according to claim 1, characterized in that, In the first stage, the global contour features of the input image are used as input to the attention mechanism, and the output is a feature representation focused on the global contour. In the second stage, the feature representation focused on the global contour is used as the input to the attention mechanism, and the output is an intermediate representation focused on the local component features; In the third stage, the intermediate representation focusing on local component features is used as the input to the attention mechanism, and the output is the target feature representation focusing on detail damage features.

6. The vehicle part identification method according to claim 1, characterized in that, After the step of outputting the target model, the following is also included: The confidence level of the recognition result obtained by the target model in recognizing the input car image is evaluated. When the confidence level of the recognition result is lower than a preset threshold, a manual review process is triggered.

7. The vehicle part identification method according to claim 6, characterized in that, Following the step of triggering the manual review process, the following steps are also included: The manually reviewed image data is added to the enhanced image dataset for subsequent incremental training.

8. A vehicle part recognition system, characterized in that, include: The image processing module is used to acquire multi-angle car images and perform image enhancement processing on the multi-angle car images to generate an enhanced image dataset; The stage division module is used to obtain the pre-trained visual model and divide the fine-tuning process of the pre-trained visual model into a first stage, a second stage, and a third stage; wherein, the pre-trained visual model includes bottom-level parameters, middle-level parameters, and top-level parameters; The first update module is used to freeze the bottom-level parameters and the middle-level parameters in the first stage, update only the top-level parameters, and perform a first round of parameter updates on the enhanced image dataset; after the first stage is completed, a first performance evaluation is performed on the trained visual model, and when the first performance evaluation result reaches a first preset performance threshold, the second stage is entered. The second update module is used to unfreeze the middle layer parameters, keep the bottom layer parameters frozen, update the middle layer parameters and the top layer parameters, and perform a second round of parameter updates on the enhanced image dataset in the second stage; after the second stage is completed, a second performance evaluation is performed on the trained visual model, and when the second performance evaluation result reaches a second preset performance threshold, the third stage is entered. The third update module is used in the third stage to unfreeze all parameters of the pre-trained visual model, update all parameters, and perform a third round of parameter updates on the augmented image dataset. The model generation module is used to output a target model after fine-tuning through the first stage, the second stage, and the third stage. The target model is used to identify car parts in the input car image.

9. An electronic device, characterized in that, The electronic device includes a processor coupled to a memory storing at least one computer program, which is loaded and executed by the processor to enable the electronic device to implement the vehicle part identification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which, when executed by a processor, implements the vehicle part identification method as described in any one of claims 1 to 7.

Images