Vehicle damage identification model establishment method, damage identification method and device

Abstract

This article provides a method for establishing a vehicle damage recognition model, a damage recognition method and an apparatus, the method comprising: determining training set data, extracting shallow image features and deep image features from the historical damage pictures; inputting the deep image features into an adaptive receptive field selection model to obtain deep robust features; fusing the deep robust features and the shallow image features to obtain target image features of the historical damage pictures; inputting the target image features into a vehicle damage recognition model to obtain image prediction and segmentation results; and adjusting the network parameters of the adaptive receptive field selection model and the vehicle damage recognition model using a preset loss function according to the image prediction and segmentation results and the damage segmentation labels corresponding to the historical damage pictures to obtain the trained adaptive receptive field selection model and the vehicle damage recognition model. This article can improve the high robustness and recognition accuracy of the training model.

Description

Technical Field

[0001] This article belongs to the field of image processing technology, and further relates to the field of vehicle insurance damage identification, specifically to a vehicle damage model establishment method, damage identification method and device. Background Art

[0002] Vehicle damage incidents often involve surface damage caused by accidents or human error, such as scratches and collisions. To facilitate vehicle repairs and insurance claims, vehicle damage assessment is necessary. Traditionally, vehicle damage assessment requires manual on-site inspections and damage assessments. However, this existing manual assessment process not only requires a significant amount of manpower but also requires highly skilled assessment personnel, resulting in a lengthy processing timeframe.

[0003] Therefore, existing technologies also utilize AI to identify vehicle damage. However, the results of this AI-based vehicle damage recognition can be biased due to various factors. For example, traditional image recognition models work better for individual targets with distinct edges, while damage in vehicle images typically lacks distinct boundaries. Furthermore, due to the influence of the shooting environment, reflections from the vehicle body, streamlined interference, and dirt often lead to false detections of damage.

[0004] Therefore, how to improve the accuracy of vehicle damage identification has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] In response to the above-mentioned problems in the prior art, the purpose of this article is to provide a vehicle damage identification model establishment method, damage identification method and device, which can improve the accuracy of vehicle damage identification.

[0006] In order to solve the above technical problems, the specific technical solutions of this article are as follows:

[0007] In one aspect, a method for establishing a vehicle damage identification model comprises:

[0008] Determining training set data, wherein the training set data includes historical damage images, the historical damage images have damage segmentation labels, and the damage segmentation labels include at least one preset damage type;

[0009] Extracting shallow image features and deep image features from the historical damage image;

[0010] Inputting the deep image features into an adaptive receptive field selection model to obtain deep robust features, wherein the adaptive receptive field selection model is used to determine different receptive field selection matrices corresponding to the deep image features according to different lesion segmentation labels;

[0011] Fusing the deep robust features and the shallow image features to obtain target image features of the historical damage image;

[0012] Inputting the target image features into a vehicle damage recognition model to obtain an image prediction segmentation result;

[0013] According to the image prediction segmentation result and the damage segmentation label corresponding to the historical damage picture, the network parameters of the adaptive receptive field selection model and the vehicle damage recognition model are adjusted using a preset loss function to obtain the trained adaptive receptive field selection model and the vehicle damage recognition model.

[0014] Furthermore, the step of inputting the deep image features into an adaptive receptive field selection model to obtain deep robust features includes:

[0015] Using a preset dilated convolution kernel, the deep image features are processed using multiple dilation rates to obtain deep image features of multiple sizes;

[0016] Fusing deep image features of multiple sizes to obtain fused deep image features;

[0017] Inputting the fused deep image features into an adaptive receptive field selection model to obtain a receptive field selection matrix, wherein the number of channels of the receptive field selection matrix is ​​consistent with the number of channels of the fused deep image features;

[0018] The deep robust features of the historical damage image are calculated based on the fused deep image features and the receptive field selection matrix.

[0019] Furthermore, the deep robust features are obtained by the following formula:

[0020] f1=multi_feature×(S+0.5×I),

[0021] Among them, f1 is the deep robust feature, multi_feature is the fused deep image feature, S is the receptive field selection matrix, and I is the unit matrix.

[0022] Furthermore, the preset loss function includes a classification loss function, an anomaly detection loss function and a receptive field selection loss function with preset weights;

[0023] The anomaly detection loss function is obtained by the following steps:

[0024] Determining the vehicle component to which the historical damage image belongs;

[0025] Determining an image feature corresponding to each vehicle component according to the vehicle component and the target image feature;

[0026] Determining a cluster center of the image features corresponding to each vehicle component based on the image features corresponding to the vehicle components;

[0027] The anomaly detection loss function is calculated based on the damage segmentation labels corresponding to the historical damage images and the cluster centers of the image features corresponding to each vehicle component.

[0028] Furthermore, the receptive field selection loss function is obtained by the following steps:

[0029] Determining a preset damage type in the damage segmentation label corresponding to the historical damage image, and the image region where the preset damage type is located;

[0030] Constructing an all-zero damage segmentation label matrix, wherein the channels of the all-zero damage segmentation label matrix correspond one-to-one to the preset damage types;

[0031] According to the image area where the preset damage type is located, the pixel values ​​of the corresponding areas on the channel corresponding to the preset damage type are sequentially set to 1 to obtain the damage segmentation label matrix corresponding to the historical damage image;

[0032] The receptive field selection loss function is calculated using a mean square error function according to the damage segmentation label matrix and the receptive field selection matrix.

[0033] On the other hand, this article also provides a vehicle damage identification method, the method comprising:

[0034] Processing the damage picture to be identified using a pre-built residual learning network to obtain shallow image features and deep image features of the damage picture to be identified;

[0035] Inputting the deep image features into the trained adaptive receptive field selection model to obtain deep robust features;

[0036] Fusing the deep robust features and the shallow image features to obtain target image features of the damage picture to be identified;

[0037] The target image features are input into the vehicle damage recognition model established by the above method to obtain the image segmentation result of the damage picture to be identified.

[0038] On the other hand, this article also provides a vehicle damage identification model establishment device, the device comprising:

[0039] A training set data determination module, configured to determine training set data, wherein the training set data includes historical damage images, the historical damage images have damage segmentation labels, and the damage segmentation labels include at least one preset damage type;

[0040] A first extraction module is used to extract shallow image features and deep image features from the historical damage image;

[0041] A first processing module is configured to input the deep image features into an adaptive receptive field selection model to obtain deep robust features, wherein the adaptive receptive field selection model is configured to determine different receptive field selection matrices corresponding to the deep image features according to different lesion segmentation labels;

[0042] A first fusion module is configured to fuse the deep robust features and the shallow image features to obtain target image features of the historical damage image;

[0043] A prediction module, configured to input the target image features into a vehicle damage recognition model to obtain an image prediction segmentation result;

[0044] A training module is used to adjust the network parameters of the adaptive receptive field selection model and the vehicle damage recognition model using a preset loss function based on the image prediction segmentation results and the damage segmentation labels corresponding to the historical damage pictures, so as to obtain the trained adaptive receptive field selection model and the vehicle damage recognition model.

[0045] On the other hand, this document also provides a vehicle damage identification device, the device comprising:

[0046] A second extraction module is used to process the damage picture to be identified using a pre-built residual learning network to obtain shallow image features and deep image features of the damage picture to be identified;

[0047] A second processing module is used to input the deep image features into the trained adaptive receptive field selection model to obtain deep robust features;

[0048] A second fusion module is used to fuse the deep robust features and the shallow image features to obtain target image features of the damage picture to be identified;

[0049] The recognition module is used to input the target image features into the vehicle damage recognition model established by the above method to obtain the image segmentation result of the damage picture to be identified.

[0050] On the other hand, this document also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described above when executing the computer program.

[0051] Finally, this document also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the method described above.

[0052] By adopting the above-mentioned technical scheme, a vehicle damage recognition model establishment method, a damage recognition method and a device described in this article, when training the vehicle damage recognition model, obtain shallow image features and deep image features in historical damage pictures, input the deep image features into the adaptive receptive field selection model to obtain deep robust features, and fuse the deep robust features and the shallow image features to obtain the target image features of the historical damage pictures; input the target image features into the vehicle damage recognition model to obtain image prediction segmentation results; according to the image prediction segmentation results and the damage segmentation labels corresponding to the historical damage pictures, use a preset loss function to adjust the network parameters of the adaptive receptive field selection model and the vehicle damage recognition model to obtain the trained adaptive receptive field selection model and the vehicle damage recognition model, which can improve the high robustness and recognition accuracy of the training model.

[0053] In order to make the above and other purposes, features and advantages of this article more obvious and easy to understand, the following specifically cites preferred embodiments and provides detailed descriptions in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of this article or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of this article. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 A schematic diagram of an implementation environment of the method provided in the embodiments of this document is shown;

[0056] Figure 2 A schematic diagram showing the steps of a method for establishing a vehicle damage identification model provided in an embodiment of this invention is shown;

[0057] Figure 3 A schematic diagram of loss segmentation labels in a historical loss image in an embodiment of this article is shown;

[0058] Figure 4 Schematic diagram of the steps for obtaining deep robust features in the embodiment of this article is shown;

[0059] Figure 5 A schematic diagram of receptive fields of different sizes corresponding to different injury types in the embodiments of this article is shown;

[0060] Figure 6 Schematic diagram showing the calculation steps of the anomaly detection loss function in the embodiment of this article;

[0061] Figure 7A schematic diagram of the calculation steps of the receptive field selection loss function in the embodiment of this article is shown;

[0062] Figure 8 A schematic diagram of the structure of a vehicle damage identification model establishment device provided in an embodiment of this invention is shown;

[0063] Figure 9 A schematic diagram showing the steps of the vehicle damage identification method provided in the embodiment of this document is shown;

[0064] Figure 10 Schematic diagram showing the calculation process of the image segmentation result in the embodiment of this article;

[0065] Figure 11 A schematic diagram of the vehicle damage identification process in the embodiment of this article is shown;

[0066] Figure 12 The following is a schematic diagram showing the structure of a vehicle damage identification device provided in an embodiment of the present invention;

[0067] Figure 13 A schematic diagram of the computer device structure provided in the embodiments of this article is shown.

[0068] Description of the accompanying symbols:

[0069] 10. Terminal;

[0070] 20. Server;

[0071] 110. Training set data determination module;

[0072] 120. First extraction module;

[0073] 130. First processing module;

[0074] 140. First fusion module;

[0075] 150. Prediction module;

[0076] 160. Training module;

[0077] 210. Second extraction module;

[0078] 220, second processing module;

[0079] 230, second fusion module;

[0080] 240, identification module;

[0081] 1302. Computer equipment;

[0082] 1304, processor;

[0083] 1306. Memory;

[0084] 1308, driving mechanism;

[0085] 1310, input / output module;

[0086] 1312. Input devices;

[0087] 1314. Output device;

[0088] 1316. Presentation equipment;

[0089] 1318. Graphical User Interface;

[0090] 1320, network interface;

[0091] 1322, communication link;

[0092] 1324. Communication bus. DETAILED DESCRIPTION

[0093] The following will be combined with the accompanying drawings to clearly and completely describe the technical solutions in the embodiments of this document. Obviously, the embodiments described are only part of the embodiments of this document, not all of the embodiments. Based on the embodiments of this document, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this document.

[0094] It should be noted that the terms "first," "second," and the like in the specification and claims herein and in the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, product, or device comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0095] Vehicle damage incidents involve identifying and confirming vehicle damage. To improve identification efficiency, existing technologies utilize artificial intelligence (AI). However, the results of AI-based vehicle damage identification can be biased due to various factors. For example, traditional image recognition models are more effective for individual targets with distinct edges, while damage in vehicle images typically lacks distinct boundaries. Furthermore, due to the influence of the shooting environment, reflections from the vehicle body, streamlined interference, and dirt often lead to false detections of damage.

[0096] In order to solve the above problems, the embodiments of this specification provide a method for establishing a vehicle damage identification model. The vehicle damage identification model established by this method can improve the accuracy and robustness of vehicle damage identification. Figure 1 As shown, it is a schematic diagram of the implementation environment of the method, which may include a terminal 10 and a server 20. A communication connection is established between the terminal 10 and the server 20 to realize data interaction. The terminal 10 can collect and store historical damage pictures, and the historical damage pictures have damage segmentation labels, and the damage segmentation labels include at least one preset damage type; that is, the terminal 10 is used to collect training set data and send the training set data to the server 20. The server 20 obtains the training set data sent by the terminal 10, and trains the adaptive receptive field selection model and the vehicle damage recognition model through the training set data, thereby obtaining a vehicle damage recognition model with high vehicle damage recognition efficiency. The specific training process can be: extracting the historical damage pictures Shallow image features and deep image features; inputting the deep image features into an adaptive receptive field selection model to obtain deep robust features, the adaptive receptive field selection model is used to determine the deep image features corresponding to different receptive field selection matrices according to different damage segmentation labels; fusing the deep robust features and the shallow image features to obtain the target image features of the historical damage picture; inputting the target image features into a vehicle damage recognition model to obtain an image prediction segmentation result; according to the image prediction segmentation result and the damage segmentation label corresponding to the historical damage picture, using a preset loss function to adjust the network parameters of the adaptive receptive field selection model and the vehicle damage recognition model to obtain the trained adaptive receptive field selection model and the vehicle damage recognition model.

[0097] In the embodiments of this specification, the server 20 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0098] In an optional embodiment, the terminal 10 may perform image segmentation processing in conjunction with the adaptive receptive field selection model and vehicle damage recognition model trained by the server 20. Specifically, the terminal 10 may include, but is not limited to, electronic devices such as smartphones, desktop computers, tablet computers, laptop computers, smart speakers, digital assistants, augmented reality (AR) / virtual reality (VR) devices, and smart wearable devices. Optionally, the operating system running on the electronic device may include, but is not limited to, Android, iOS, Linux, and Windows.

[0099] In addition, it should be noted that Figure 1 What is shown is only one application environment provided by the present disclosure. In actual applications, other application environments may also be included, such as the training of an adaptive receptive field selection model and a vehicle damage recognition model, which may also be implemented on the terminal 10.

[0100] Specifically, the embodiments of this article provide a method for establishing a vehicle damage identification model, which can improve the accuracy of vehicle damage identification. Figure 2 It is a schematic diagram of the steps of a method for establishing a vehicle damage identification model provided in the embodiment of this article. This specification provides the method operation steps as described in the embodiment or flowchart, but it may include more or fewer operation steps based on conventional or non-creative labor. The order of steps listed in the embodiment is only one way of executing the steps among many steps, and does not represent the only execution order. When the actual system or device product is executed, it can be executed in the order of the method shown in the embodiment or the accompanying drawings or in parallel. Specifically, Figure 2 As shown, the method may include:

[0101] S201: Determine training set data, where the training set data includes historical damage images, the historical damage images have damage segmentation labels, and the damage segmentation labels include at least one preset damage type;

[0102] S202: Extracting shallow image features and deep image features from the historical damage image;

[0103] S203: Inputting the deep image features into an adaptive receptive field selection model to obtain deep robust features, wherein the adaptive receptive field selection model is used to determine different receptive field selection matrices corresponding to the deep image features according to different damage segmentation labels;

[0104] S204: fusing the deep robust features and the shallow image features to obtain target image features of the historical damage image;

[0105] S205: Inputting the target image features into a vehicle damage recognition model to obtain an image prediction segmentation result;

[0106] S206: Based on the image prediction segmentation result and the damage segmentation label corresponding to the historical damage image, a preset loss function is used to adjust the network parameters of the adaptive receptive field selection model and the vehicle damage recognition model to obtain the trained adaptive receptive field selection model and the vehicle damage recognition model.

[0107] It can be understood that the execution entity of the embodiment of this specification can be a server. When training the vehicle damage recognition model, by adding an adaptive receptive field selection model to process deep image features, different receptive fields can be selected for different damage types, thereby obtaining image features of appropriate sizes, and then fusing them with shallow features of the image to obtain vehicle damage features (i.e., target image features) that have both robustness and appropriate size information. Training the vehicle damage recognition model through the vehicle damage features can improve the accuracy of model training and can adapt to complex external environments.

[0108] The shallow image features may be visual collection features of historical damage pictures, such as texture information, shape information, and color information in the pictures, and the deep image features may be semantic information in historical damage pictures, such as the deep image features representing different damage types.

[0109] The damage segmentation label can be the area occupied by different damage types in the same damage image. That is, there may be multiple different damage types in the same damage image, such as Figure 3 As shown in the figure, in a damage image, area A1 is the deformed area and area A2 is the scratched area. By dividing the damage segmentation labels in the damage image, corresponding image features can be extracted for different areas. The adaptive receptive field selection model selects different receptive fields according to different deep image features, thereby improving the accuracy of damage image feature extraction and the accuracy and robustness of model training.

[0110] The preset damage types include at least abrasions, deformations, ruptures and breakages, and may also include other damage types. All types of the preset damage types are not limited in the embodiments of this specification.

[0111] In the embodiment of this specification, the extraction of shallow image features and deep image features from the historical damage image includes:

[0112] The historical damage images are processed using a pre-built residual learning network to obtain shallow image features and deep image features.

[0113] For example, a backbone network with the ResNet50 network structure as the main body can be built, and the fully connected network layer in its structure can be removed. When inputting a vehicle image with a size of W×H×3, the input image is first cropped to meet the preset input size, and then the cropped image is processed and calculated using the backbone network built in advance. The output features of the block_1 layer in the network structure are recorded as the shallow features of the image, and the output features of the block_4 layer in the network structure are recorded as the deep features of the image.

[0114] In the embodiments of this specification, Figure 4 As shown, the deep image features are input into the adaptive receptive field selection model to obtain deep robust features, including:

[0115] S401: using a preset dilated convolution kernel and multiple dilation rates to process the deep image features to obtain deep image features of multiple sizes;

[0116] S402: Fusing deep image features of multiple sizes to obtain fused deep image features;

[0117] S403: Inputting the fused deep image features into an adaptive receptive field selection model to obtain a receptive field selection matrix, wherein the number of channels of the receptive field selection matrix is ​​consistent with the number of channels of the fused deep image features;

[0118] S404: Calculate and obtain deep robust features of the historical damage image based on the fused deep image features and the receptive field selection matrix.

[0119] It can be understood that by processing deep image features with preset hole convolution kernels, image features of receptive fields of different sizes can be obtained, ensuring a certain image resolution. At the same time, there is no need to increase additional computing resources. Using multiple expansion rates can capture multi-scale image information.

[0120] Since the most suitable receptive fields for different vehicle damages are different (i.e. different damage types occupy different sizes, for example, the size occupied by deformation is larger than that occupied by cracks), the receptive field used for vehicle deformation generally needs to be larger, otherwise it is difficult to see the exact characteristics of the deformation. Vehicle damage features often have some small cracks, so using a larger receptive field convolution will cause the target features to account for a smaller proportion, resulting in the network being unable to learn these features, resulting in reduced model reliability. For example, Figure 5As shown, B1 represents the damaged area, and B2 represents the scratched area. The appropriate receptive fields required for B1 and B2 are significantly different. The adaptive receptive field selection model generates a receptive field selection matrix that adaptively selects image features of corresponding sizes based on different preset damage types. The number of channels in this receptive field selection matrix matches the number of channels in the fused deep image features, allowing for more accurate deep image features to be determined based on the receptive field selection matrix.

[0121] In the embodiment of this specification, before fusing the deep image features of multiple sizes to obtain the fused deep image features, the following steps may be further included:

[0122] Determining a deep image feature of a maximum size from the deep image features of the plurality of sizes;

[0123] According to the deep image feature of the maximum size, the deep image features of other sizes are expanded to the maximum size, so that all the deep image features are unified in the same size;

[0124] The deep image features after the unified size are fused to obtain the fused deep image features.

[0125] After unifying the sizes of deep image features of multiple sizes, it is convenient for feature fusion processing. For example, the deep image features of multiple sizes are 3X3X1, 5X5X1, and 8X8X1, respectively. Then 8X8X1 can be considered as the deep image feature of the largest size, and then 3X3X1 and 5X5X1 are both expanded to the size of 8X8. In specific implementation, the upsampling layer can be used to unify the sizes of the deep image features of the above multiple sizes to wk×hk×ck. The fusion processing can connect the channels of multiple features of the same size (i.e., channel fusion).

[0126] The adaptive receptive field selection model includes a convolution layer and an activation layer.

[0127] Furthermore, the fused deep image features are input into an adaptive receptive field selection model to obtain a receptive field selection matrix, including:

[0128] Inputting the fused deep image features into the convolution layer for convolution processing to obtain feature maps of deep feature images of different sizes corresponding to different preset damage types in the fused deep image features;

[0129] The feature map is input into the activation layer for normalization processing to obtain a receptive field selection matrix, and the number of channels of the receptive field selection matrix is ​​consistent with the number of feature channels of the fused deep image.

[0130] It should be noted that, in addition to the convolution layer and the activation layer, the adaptive receptive field selection model may also include the preset hole convolution kernel for extracting deep image features of multiple sizes. Accordingly, the adaptive receptive field selection model may include a preset hole convolution layer, a convolution layer and an activation layer. The functions of the preset hole convolution layer, the convolution layer and the activation layer are consistent with the above method and will not be elaborated in the embodiments of this specification.

[0131] It should be noted again that the adaptive receptive field selection model can assign different receptive fields (i.e., different receptive field convolution kernels) to different damage types, but the adaptive receptive field selection model needs to be trained to achieve the corresponding effect. Therefore, it is also necessary to set the receptive field selection matrix label in the historical damage images in the training set data to facilitate the training of the adaptive receptive field selection model. The trained model can ensure that the final features have appropriate size (or scale) information, which is more accurate than the full scale information used by the traditional feature pyramid network (FPN). There is no redundant scale information in the features. On the other hand, the use of this adaptive receptive field selection network only requires 4 different void convolution kernel parameters (i.e., void convolution kernel parameters consistent with the number of preset damage types) and 1 layer of convolution network parameters. The use of stacked convolutional neural networks such as FPN will introduce a huge amount of feature calculation and network parameters, which poses a great challenge to the timeliness and storage space of image calculation, and is not conducive to the realization of lightweight model front-end functions.

[0132] For example, taking the four preset damage types of abrasion, deformation, rupture and breakage as an example, on the basis of obtaining the deep image features, the deep features of the image are processed using a dilated convolution kernel with a kernel size of 3x3 and rate sizes of 6, 12, 18 and 24 respectively, and four image features of different sizes are obtained. Then, the upsampling layer is used to unify the feature sizes of these image features into wk×hk×ck.

[0133] For the adaptive receptive field selection model, a conventional convolutional layer is used to process deep image features and its value is normalized to between 0 and 1 using the Sigmoid function, and the receptive field selection matrix S∈R is output. wk×hk×ck×4 , and then the deep robust features of the historical damage image are calculated using the following formula (1):

[0134] f1=multi_feature×(S+0.5×I), (1)

[0135] Among them, f1 is the deep robust feature, multi_feature is the fused deep image feature, S is the receptive field selection matrix, I is the unit matrix, and the unit matrix has the same size as the receptive field selection matrix.

[0136] In some other embodiments, the deep robust features may also be obtained by other calculation methods, which are not limited in the embodiments of this specification.

[0137] In addition to processing deep image features, the embodiment of this specification may also process shallow image features. Optionally, before fusing the deep robust features and the shallow image features to obtain the target image features of the historical damage image, the process further includes:

[0138] The shallow image features are subjected to feature enhancement processing using a feature enhancement network to obtain feature-enhanced shallow image features.

[0139] Generally speaking, the difference between the damaged part of a vehicle and the normal parts around it will be reflected in texture and color. The regional features of the damaged part can be enhanced through the feature enhancement network, which increases the difference from the surrounding normal parts and improves the accuracy of model training. For example, the shallow image features can be processed through the SE_block network structure with a dimension of 256, so as to enhance the regional features of the damaged part in the historical damage image.

[0140] After processing the deep image features and shallow image features, the processed features can be fused, that is, the scale-robust deep image features and the shallow image features that enhance texture and color information are channel-fused, so that the fused features used to predict the segmentation results have rich scale features, thereby obtaining more reliable training results during model training, and effectively distinguishing the differences between the damaged area and the surrounding area.

[0141] In an embodiment of the present specification, the vehicle damage recognition model can be used for preset damage types (such as n types of damage types). By receiving the fused target image features, the output matrix size should be W×H×n, where W is the image width and H is the image height. It can be regarded as n probability maps of the same size as the original input image, and each probability map corresponds to a damage type. The value of a pixel position on the i-th probability map represents the probability value of the pixel at the same position in the original input image belonging to the i-th type of damage. The point with the largest probability value of the pixel at the same position in each probability map is counted as the target pixel. In this way, a target probability map can be obtained. The distribution of pixels in the target probability map (different pixels correspond to different damage types) is the vehicle damage prediction result of the model. Exemplarily, the structure of the vehicle damage recognition model can be to stack 2 layers of deconvolution networks and 1 layer of upsampling layer to achieve the above function and obtain the vehicle damage segmentation probability result.

[0142] In the embodiment of this specification, the preset loss function includes a classification loss function, an anomaly detection loss function, and a receptive field selection loss function with preset weights. Alternatively, the preset damage function can be obtained by the following formula (2):

[0143] L=λ1L cls +λ2L dis +λ3L sel , (2)

[0144] Among them, L cls is the classification loss of the segmentation result, L dis is the anomaly detection loss function of the fusion feature, L sel Select the loss function for the receptive field, λ1, λ2 and λ3 represent the weights of the above loss function, λ1+λ2+λ3=1.

[0145] The classification loss function can be a loss function calculated based on the preset image segmentation results and the damage segmentation label. Optionally, it can be a cross-entropy loss function of multiple binary classifications. The specific calculation process is the conventional calculation process of the loss function, which is not described in detail in the embodiments of this specification. Of course, the classification loss function can also be other types of loss functions, which are not limited in this specification.

[0146] The anomaly detection loss function is the inter-class distance of different fusion features in the historical damage image, such as Figure 6 As shown, the calculation process is as follows:

[0147] S601: Determine the vehicle component to which the historical damage image belongs;

[0148] S602: Determine an image feature corresponding to each vehicle component based on the vehicle component and the target image feature;

[0149] S603: Determine the cluster center of the image features corresponding to each vehicle component based on the image features corresponding to the vehicle components;

[0150] S604: Calculate the anomaly detection loss function based on the damage segmentation labels corresponding to the historical damage images and the cluster centers of the image features corresponding to each vehicle component.

[0151] It can be understood that, unlike traditional anomaly detection models that use abnormal areas as units, traditional anomaly detection models use feature clustering or maximize class distances, the main purpose of which is to distinguish damaged areas from surrounding areas. The anomaly detection loss function in this article is based on different parts of the vehicle, segmenting the features of different parts, calculating the inter-class distance between image features of different damaged parts, and maximizing the difference between normal vehicle parts and abnormal vehicle parts, effectively solving the problem of misdetection of the streamline shape of the vehicle body and improving the detection accuracy of damaged parts.

[0152] Alternatively, the anomaly detection loss function is obtained by the following formula (3):

[0153]

[0154] in, represents the Euclidean distance between the cluster centers of the image features corresponding to the i-th and j-th vehicle parts, represents the cluster center of the image feature corresponding to the i-th type of vehicle parts, M represents the total number of pixels in the image feature corresponding to the i-th type of vehicle parts, K i is the pixel point set of the image feature corresponding to the i-th type of vehicle parts, p is the pixel coordinate in the image feature corresponding to the i-th type of vehicle parts, δ is a priori value, for example, it can be taken as 1.5, Nor represents the set of undamaged vehicle parts, and AN represents the set of damaged vehicle parts.

[0155] In the actual training process, it is necessary to label the vehicle parts belonging to the historical damage pictures in the training set data in advance. In this way, when calculating the anomaly detection loss function, the component label information in the historical damage picture can be directly input. Combined with the above formula (3), the value of the calculated loss function can be obtained. It can be clearly seen from the above formula (3) that when the distance between the feature clustering center of the damaged vehicle parts and the normal vehicle parts increases, the value of the anomaly detection loss function will decrease accordingly. Therefore, using this anomaly detection loss function to assist training can maximize the difference between damaged and undamaged parts, so that the network can implicitly reduce the situation of false detection of damage caused by the streamlined structure of the vehicle itself during the learning and training process.

[0156] In the process of receptive field selection processing for deep image features, it is necessary to obtain a receptive field selection matrix, which is the output result of the adaptive receptive field selection model. In order to make the adaptive receptive field selection model converge, it is necessary to further set the receptive field selection matrix label (i.e., the damage segmentation label matrix) for the damage segmentation label corresponding to the historical damage image as the target result output by the model, so that the loss function of the adaptive receptive field selection model can be calculated by the receptive field selection matrix label and the receptive field selection matrix, and then the model can be trained. As an option, Figure 7 As shown, the receptive field selection loss function is obtained by the following steps:

[0157] S701: Determine a preset damage type in a damage segmentation label corresponding to the historical damage image, and the image region where the preset damage type is located;

[0158] S702: Constructing an all-zero damage segmentation label matrix, wherein the channels of the all-zero damage segmentation label matrix correspond one-to-one to the preset damage types;

[0159] S703: According to the image area where the preset damage type is located, pixel values ​​of corresponding areas on the channel corresponding to the preset damage type are sequentially set to 1 to obtain a damage segmentation label matrix corresponding to the historical damage image;

[0160] S704: According to the damage segmentation label matrix and the receptive field selection matrix, a mean square error function is used to calculate the receptive field selection loss function.

[0161] It can be understood that the damage segmentation label matrix is ​​the target result of inputting deep image features into the adaptive receptive field selection model. By setting the number of channels consistent with the number of preset damage types, for example, if there are four damage types, the damage segmentation label matrix has four channels, each channel corresponds to a damage type, and the channels are set in sequence according to the adaptive scale of the damage type. The first channel corresponds to the damage type with the smallest required scale (or size), and the fourth channel corresponds to the damage type with the largest required scale. By setting corresponding receptive field selection matrix labels for different damage types, the adaptive receptive field selection model is prompted to assign different receptive field convolution kernels to different damage types, thereby ensuring that there is appropriate scale information in the features and improving the reliability and accuracy of feature extraction.

[0162] For example, if the damage segmentation label matrix of the vehicle damage recognition model is Label, then the region with pixel value j in the label matrix represents the region corresponding to the jth type of vehicle damage. First, an all-zero matrix Z of size wk×hk×4 is generated. Then, all pixel locations of the deformation type regions in all Labels are searched and the pixel values ​​of these locations are set to 1 on the fourth channel of Z. The same location search is performed for the damaged type regions, and the pixel values ​​of these locations are set to 1 on the first channel of Z. The corresponding operation is also performed on the second and third channels. The resulting damage segmentation label matrix can make the receptive field selection tendency corresponding to different damages more clear.

[0163] The calculation result of each training of the preset loss function can be obtained by calculating the damage function, thereby realizing the training of the damage adaptive receptive field selection model and the vehicle damage recognition model. Optionally, the network parameters of the adaptive receptive field selection model and the vehicle damage recognition model are adjusted using the preset loss function based on the image prediction segmentation result and the damage segmentation label corresponding to the historical damage image, to obtain the trained adaptive receptive field selection model and the vehicle damage recognition model, including:

[0164] According to the image prediction segmentation result and the damage segmentation label corresponding to the historical damage image, a loss function calculation value is calculated using the preset loss function;

[0165] Determine whether the calculated value of the loss function is lower than a loss threshold;

[0166] If not, adjusting the network parameters of the adaptive receptive field selection model and the vehicle damage recognition model, and returning to the shallow image feature extraction step and the deep image feature extraction step until the calculated loss function value is lower than the loss threshold, and obtaining the final network parameters of the adaptive receptive field selection model and the vehicle damage recognition model;

[0167] If so, obtaining the final network parameters of the adaptive receptive field selection model and the vehicle damage identification model;

[0168] According to the final network parameters of the adaptive receptive field selection model and the vehicle damage recognition model, the trained adaptive receptive field selection model and the vehicle damage recognition model are obtained.

[0169] It can be understood that in the embodiments of this specification, the training set data is feature processed through an adaptive receptive field selection model, and the processing results are used as the input results of the vehicle damage recognition model, and then the image prediction segmentation results are obtained and trained by the loss function calculated by the above formula. The loss threshold can be set according to actual conditions and is not limited in the embodiments of this specification.

[0170] During the training process, the network parameters of the adaptive receptive field selection model and the vehicle damage identification model can be adjusted by backpropagating the calculated loss function value until the two models reach convergence.

[0171] The vehicle damage recognition model establishment method provided by the embodiment of this specification extracts shallow image features and deep image features of the image during the training process, learns the semantic features of vehicle damage from the deep features of the image, and uses different scales of void convolution to process the semantic features. Then, according to the adaptive receptive field selection model, the receptive field convolution results of the images of different scales are adaptively selected. Finally, the deep semantic features with appropriate scale information are channel-fused with the enhanced shallow visual features. The fused features are used to predict and output the segmentation results of the damaged parts of the vehicle body. Furthermore, the abnormal detection damage function used by the vehicle damage recognition model can effectively compare normal and abnormal areas of the vehicle parts, so that the network can learn the difference between suspected damage and real damage caused by complex environments, greatly alleviating the misjudgment of damage recognition. In addition, the adaptive receptive field selection model is adjusted according to the feedback of the damage segmentation results, which can more effectively use the optimal receptive field scale for different damage types. Through the above model construction, the vehicle damage recognition model has high robustness and recognition accuracy for the shooting environment of the vehicle image and common damage types.

[0172] Based on the same inventive concept, the embodiment of this specification also provides a vehicle damage identification model establishment device, such as Figure 8 As shown, the device includes:

[0173] A training set data determination module 110 is configured to determine training set data, wherein the training set data includes historical damage images, the historical damage images have damage segmentation labels, and the damage segmentation labels include at least one preset damage type;

[0174] A first extraction module 120 is configured to extract shallow image features and deep image features from the historical damage image;

[0175] A first processing module 130 is configured to input the deep image features into an adaptive receptive field selection model to obtain deep robust features, wherein the adaptive receptive field selection model is configured to determine different receptive field selection matrices corresponding to the deep image features according to different lesion segmentation labels;

[0176] A first fusion module 140 is configured to fuse the deep robust features and the shallow image features to obtain target image features of the historical damage image;

[0177] Prediction module 150, used to input the target image features into the vehicle damage recognition model to obtain image prediction segmentation results;

[0178] The training module 160 is used to adjust the network parameters of the adaptive receptive field selection model and the vehicle damage recognition model based on the image prediction segmentation results and the damage segmentation labels corresponding to the historical damage pictures using a preset loss function to obtain the trained adaptive receptive field selection model and the vehicle damage recognition model.

[0179] The beneficial effects achieved by the above-mentioned device are consistent with the beneficial effects achieved by the above-mentioned method, and are not described in detail in the embodiments of this specification.

[0180] Based on the above-mentioned method for establishing a vehicle damage identification model, the embodiment of this specification also provides a vehicle damage identification method, such as Figure 9 As shown, the method includes:

[0181] S901: Processing the damage picture to be identified using a pre-built residual learning network to obtain shallow image features and deep image features of the damage picture to be identified;

[0182] S902: Inputting the deep image features into the trained adaptive receptive field selection model to obtain deep robust features;

[0183] S903: Fusing the deep robust features and the shallow image features to obtain target image features of the damage picture to be identified;

[0184] S904: Inputting the target image features into the vehicle damage recognition model established by the above method to obtain an image segmentation result of the damage picture to be recognized.

[0185] Furthermore, if Figure 10 As shown, the target image features are input into the trained vehicle damage recognition model to obtain the image segmentation result of the damage picture to be identified, including:

[0186] S1001: Inputting the target image features into a trained vehicle damage recognition model to obtain an identification type distribution matrix, wherein the channel output probability maps of the identification type distribution matrix correspond one-to-one to preset damage types, and the value of each pixel in the channel output probability map represents the predicted probability of the pixel in the channel output probability map;

[0187] S1002: traversing the predicted probabilities of pixels at the same position in different channel output probability maps, determining the maximum predicted probability of pixels at the same position, and obtaining a recognition probability map of the damage image to be recognized;

[0188] S1003: Determine the distribution of different preset damage types according to the recognition probability map;

[0189] S1004: Determine the image segmentation result of the damage picture to be identified according to the distribution of the different preset damage types.

[0190] For example, Figure 11 The figure shows a schematic diagram of the vehicle damage recognition process. The image to be identified is input into the backbone network with the ResNet50 network structure as the main body to obtain shallow image features and deep image features. The shallow image features are input into the SE enhancement network for feature enhancement processing. The deep image features are input into the adaptive receptive field selection network to obtain deep image features of appropriate scale. Feature fusion processing is then performed to obtain fused features. The fused features are then input into the trained vehicle damage recognition model to obtain the segmentation results of the image to be identified. The model provided in this article can improve the accuracy of vehicle damage recognition.

[0191] Based on the same inventive concept, this article also provides a vehicle damage identification device, such as Figure 12 As shown, the device includes:

[0192] The second extraction module 210 is used to process the damage picture to be identified using a pre-built residual learning network to obtain shallow image features and deep image features of the damage picture to be identified;

[0193] A second processing module 220 is configured to input the deep image features into the trained adaptive receptive field selection model to obtain deep robust features;

[0194] A second fusion module 230 is configured to fuse the deep robust features and the shallow image features to obtain target image features of the damage image to be identified;

[0195] The recognition module 240 is used to input the target image features into the vehicle damage recognition model established by the above method to obtain the image segmentation result of the damage picture to be recognized.

[0196] like Figure 13As shown, a computer device provided in an embodiment of this document is provided. The apparatus herein may be a computer device in this embodiment, executing the method described above. The computer device 1302 may include one or more processors 1304, such as one or more central processing units (CPUs), each of which may implement one or more hardware threads. The computer device 1302 may also include any memory 1306 for storing any type of information, such as code, settings, data, etc. For example, without limitation, the memory 1306 may include any one or more combinations of the following: any type of RAM, any type of ROM, a flash memory device, a hard disk, an optical disk, etc. More generally, any memory may use any technology to store information. Furthermore, any memory may provide volatile or non-volatile retention of information. Furthermore, any memory may represent a fixed or removable component of the computer device 1302. In one embodiment, when the processor 1304 executes associated instructions stored in any memory or combination of memories, the computer device 1302 may perform any operation of the associated instructions. The computer device 1302 also includes one or more drive mechanisms 1308 for interacting with any storage, such as a hard disk drive mechanism, an optical disk drive mechanism, and the like.

[0197] Computer device 1302 may also include an input / output module 1310 (I / O) for receiving various inputs (via input devices 1312) and providing various outputs (via output devices 1314). A specific output mechanism may include a presentation device 1316 and an associated graphical user interface (GUI) 1318. In other embodiments, input / output module 1310 (I / O), input devices 1312, and output devices 1314 may not be included, and the computer device 1302 may simply be a computer device on a network. Computer device 1302 may also include one or more network interfaces 1320 for exchanging data with other devices via one or more communication links 1322. One or more communication buses 1324 couple the components described above together.

[0198] The communication link 1322 can be implemented in any manner, for example, through a local area network, a wide area network (e.g., the Internet), a point-to-point connection, etc., or any combination thereof. The communication link 1322 can include any combination of hardwired links, wireless links, routers, gateway functions, name servers, etc., governed by any protocol or combination of protocols.

[0199] Corresponding to Figure 2 、 Figure 4 、 Figure 6-Figure 7 、 Figure 9-10The embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which executes the steps of the above method when executed by a processor.

[0200] The embodiment of the present invention also provides a computer readable instruction, wherein when the processor executes the instruction, the program causes the processor to execute the following Figure 2 、 Figure 4 、 Figure 6-Figure 7 、 Figure 9-10 The method shown in .

[0201] It should be understood that in the various embodiments of this document, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this document.

[0202] It should also be understood that in the embodiments herein, the term "and / or" merely describes an association between associated objects, indicating that three possible relationships exist. For example, "A and / or B" could represent: A alone, A and B simultaneously, or B alone. Furthermore, the character " / " in this document generally indicates an "or" relationship between the associated objects.

[0203] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this document.

[0204] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0205] In the several embodiments provided herein, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices, or units, or can be an electrical, mechanical, or other form of connection.

[0206] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiments herein.

[0207] In addition, the functional units in the various embodiments herein may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0208] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this article is essentially or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of this article. The aforementioned storage medium includes: various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0209] This article uses specific embodiments to illustrate the principles and implementation methods of this article. The description of the above embodiments is only used to help understand the methods and core ideas of this article. At the same time, for those skilled in the art, based on the ideas of this article, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation to this article.

Claims

1. A method for establishing a vehicle damage identification model, characterized in that: The method comprises: Determining training set data, wherein the training set data includes historical damage images, the historical damage images have damage segmentation labels, and the damage segmentation labels include at least one preset damage type; Extracting shallow image features and deep image features from the historical damage image; Inputting the deep image features into an adaptive receptive field selection model to obtain deep robust features, wherein the adaptive receptive field selection model is used to determine different receptive field selection matrices corresponding to the deep image features according to different lesion segmentation labels; Fusing the deep robust features and the shallow image features to obtain target image features of the historical damage image; Inputting the target image features into a vehicle damage recognition model to obtain an image prediction segmentation result; According to the image prediction segmentation result and the damage segmentation label corresponding to the historical damage image, a preset loss function is used to adjust the network parameters of the adaptive receptive field selection model and the vehicle damage recognition model to obtain the trained adaptive receptive field selection model and the vehicle damage recognition model; The preset loss function includes a classification loss function, an anomaly detection loss function and a receptive field selection loss function with preset weights; The anomaly detection loss function is obtained by the following steps: Determining the vehicle component to which the historical damage image belongs; Determining an image feature corresponding to each vehicle component according to the vehicle component and the target image feature; Determining a cluster center of the image features corresponding to each vehicle component based on the image features corresponding to the vehicle components; The anomaly detection loss function is calculated based on the damage segmentation labels corresponding to the historical damage images and the cluster centers of the image features corresponding to each vehicle component.

2. The method according to claim 1, characterized in that Inputting the deep image features into an adaptive receptive field selection model to obtain deep robust features includes: Using a preset dilated convolution kernel, the deep image features are processed using multiple dilation rates to obtain deep image features of multiple sizes; Fusing deep image features of multiple sizes to obtain fused deep image features; Inputting the fused deep image features into an adaptive receptive field selection model to obtain a receptive field selection matrix, wherein the number of channels of the receptive field selection matrix is ​​consistent with the number of channels of the fused deep image features; The deep robust features of the historical damage image are calculated based on the fused deep image features and the receptive field selection matrix.

3. The method according to claim 2, characterized in that The deep robust features are obtained by the following formula: f1=multi_feature×(S+0.5×I), Among them, f1 is the deep robust feature, multi_feature is the fused deep image feature, S is the receptive field selection matrix, and I is the unit matrix.

4. The method according to claim 1, wherein The receptive field selection loss function is obtained by the following steps: Determining a preset damage type in the damage segmentation label corresponding to the historical damage image, and the image region where the preset damage type is located; Constructing an all-zero damage segmentation label matrix, wherein the channels of the all-zero damage segmentation label matrix correspond one-to-one to the preset damage types; According to the image area where the preset damage type is located, the pixel values ​​of the corresponding areas on the channel corresponding to the preset damage type are sequentially set to 1 to obtain the damage segmentation label matrix corresponding to the historical damage image; The receptive field selection loss function is calculated using a mean square error function according to the damage segmentation label matrix and the receptive field selection matrix.

5. A vehicle damage identification method, characterized in that: The method comprises: Processing the damage picture to be identified using a pre-built residual learning network to obtain shallow image features and deep image features of the damage picture to be identified; Inputting the deep image features into the trained adaptive receptive field selection model to obtain deep robust features; Fusing the deep robust features and the shallow image features to obtain target image features of the damage picture to be identified; The target image features are input into a vehicle damage recognition model established by the method of any one of claims 1 to 4 to obtain an image segmentation result of the damage picture to be identified.

6. A vehicle damage identification model establishment device, characterized in that: The device comprises: A training set data determination module, configured to determine training set data, wherein the training set data includes historical damage images, the historical damage images have damage segmentation labels, and the damage segmentation labels include at least one preset damage type; A first extraction module is used to extract shallow image features and deep image features from the historical damage image; A first processing module is configured to input the deep image features into an adaptive receptive field selection model to obtain deep robust features, wherein the adaptive receptive field selection model is configured to determine different receptive field selection matrices corresponding to the deep image features according to different lesion segmentation labels; A first fusion module is configured to fuse the deep robust features and the shallow image features to obtain target image features of the historical damage image; A prediction module, configured to input the target image features into a vehicle damage recognition model to obtain an image prediction segmentation result; a training module, configured to adjust network parameters of the adaptive receptive field selection model and the vehicle damage recognition model using a preset loss function based on the image prediction segmentation result and the damage segmentation label corresponding to the historical damage image, thereby obtaining the trained adaptive receptive field selection model and the vehicle damage recognition model; The preset loss function includes a classification loss function, an anomaly detection loss function and a receptive field selection loss function with preset weights; The anomaly detection loss function is obtained by the following steps: Determining the vehicle component to which the historical damage image belongs; Determining an image feature corresponding to each vehicle component according to the vehicle component and the target image feature; Determining a cluster center of the image features corresponding to each vehicle component based on the image features corresponding to the vehicle components; The anomaly detection loss function is calculated based on the damage segmentation labels corresponding to the historical damage images and the cluster centers of the image features corresponding to each vehicle component.

7. A vehicle damage identification device, characterized in that: The device comprises: A second extraction module is used to process the damage picture to be identified using a pre-built residual learning network to obtain shallow image features and deep image features of the damage picture to be identified; A second processing module is used to input the deep image features into the trained adaptive receptive field selection model to obtain deep robust features; A second fusion module is used to fuse the deep robust features and the shallow image features to obtain target image features of the damage picture to be identified; The recognition module is used to input the target image features into the vehicle damage recognition model established by the method of any one of claims 1 to 4 to obtain the image segmentation result of the damage picture to be identified.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 5 is implemented.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images