Vehicle damage detection method, training method, device, equipment and storage medium

Abstract

The invention relates to the technical field of vehicle damage detection, and provides a vehicle damage detection method and device, a training method and device, equipment and a storage medium, and the method comprises the steps: obtaining a vehicle image; performing feature extraction processing on the vehicle image based on a backbone network of a vehicle damage detection model to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network comprises a deep separable convolutional layer; the size of a convolution kernel in the depth separable convolution layer is greater than or equal to a preset size threshold value; performing feature enhancement and fusion processing on the feature information based on a neck network of a vehicle damage detection model to obtain processed feature information; the network architecture of the neck network is the same as that of the backbone network; and determining vehicle damage information corresponding to the vehicle image according to the processed feature information based on a head network of a vehicle damage detection model, so as to improve the detection accuracy of the vehicle damage information and improve the detection speed of the vehicle damage information.

Description

Technical Field

[0001] The present application relates to the technical field of vehicle damage detection, and in particular to a vehicle damage detection method, training method, device, equipment and storage medium. Background Art

[0002] With the development of object detection, object detection models can be applied to automatic vehicle damage assessment to detect vehicle damage. The YOLO (You Only Look Once) series improves the accuracy and efficiency of single-stage object detectors by exploring different model architectures and training strategies.

[0003] In the related art, it is generally chosen to use a larger backbone network with more parameters, or a heavier neck network, or to make more connections and fusions between feature pyramids to enhance the feature information of the detection target, so as to improve the detection accuracy of the detection target. However, these attempts have increased the calculation and memory usage of the target detection model, which can easily lead to poor detection speed of the detection target. For example, for vehicle damage information, if the vehicle damage detection model has a poor vehicle damage detection speed, it is easy to cause users to have a poor experience with the vehicle damage detection model. Summary of the invention

[0004] The main purpose of this application is to provide a vehicle damage detection method, training method, device, equipment and storage medium, aiming to improve the detection accuracy of vehicle damage information while improving the detection speed of vehicle damage information.

[0005] In a first aspect, the present application provides a vehicle damage detection method, comprising:

[0006] Acquire vehicle images;

[0007] Based on the backbone network of the vehicle damage detection model, feature extraction processing is performed on the vehicle image to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depthwise separable convolution layer; the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to a preset size threshold;

[0008] Based on the neck network of the vehicle damage detection model, feature enhancement and fusion processing are performed on the feature information to obtain processed feature information; the network architecture of the neck network is the same as the network architecture of the backbone network;

[0009] Based on the head network of the vehicle damage detection model, the vehicle damage information corresponding to the vehicle image is determined according to the processed feature information.

[0010] In a second aspect, the present application also provides a method for training a vehicle damage detection model, comprising:

[0011] Acquire a training sample set, wherein the training sample set includes a plurality of vehicle images and vehicle damage information labels corresponding to the plurality of vehicle images;

[0012] Based on the backbone network of the vehicle damage detection model, feature extraction processing is performed on the vehicle image to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depthwise separable convolution layer; the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to a preset size threshold;

[0013] Based on the neck network of the vehicle damage detection model, feature enhancement and fusion processing are performed on the feature information to obtain processed feature information; the network architecture of the neck network is the same as the network architecture of the backbone network;

[0014] Based on the head network of the vehicle damage detection model, determining the vehicle damage information corresponding to the vehicle image according to the processed feature information;

[0015] Determining a model optimization parameter according to an intersection-and-union ratio between the vehicle damage information and the vehicle damage information label;

[0016] According to the model optimization parameters, the model parameters of the vehicle damage detection model are adjusted.

[0017] In a third aspect, the present application further provides a vehicle damage detection device, the vehicle damage detection device comprising:

[0018] An image acquisition module, used for acquiring vehicle images;

[0019] A feature extraction module is used to perform feature extraction processing on the vehicle image based on the backbone network of the vehicle damage detection model to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depthwise separable convolution layer; the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to a preset size threshold;

[0020] A feature processing module, used for performing feature enhancement and fusion processing on the feature information based on the neck network of the vehicle damage detection model to obtain processed feature information; the network architecture of the neck network is the same as the network architecture of the backbone network;

[0021] A vehicle damage recognition module is used to determine the vehicle damage information corresponding to the vehicle image based on the head network of the vehicle damage detection model and according to the processed feature information.

[0022] In a fourth aspect, the present application further provides a computer device, the computer device comprising a memory and a processor;

[0023] The memory is used to store computer programs;

[0024] The processor is used to execute the computer program and implement the steps of the vehicle damage detection method as described above when executing the computer program, or implement the steps of the vehicle damage detection model training method as described above.

[0025] In a fifth aspect, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the vehicle damage detection method as described above are implemented, or the steps of the vehicle damage detection model training method as described above are implemented.

[0026] The present application provides a vehicle damage detection method, training method, device, equipment and storage medium. The vehicle damage detection method includes: obtaining a vehicle image; based on the backbone network of the vehicle damage detection model, performing feature extraction processing on the vehicle image to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depth-separable convolution layer; the size of the convolution kernel in the depth-separable convolution layer is greater than or equal to a preset size threshold; based on the neck network of the vehicle damage detection model, feature enhancement and fusion processing are performed on the feature information to obtain processed feature information; the network architecture of the neck network is the same as the network architecture of the backbone network; based on the head network of the vehicle damage detection model, the vehicle damage information corresponding to the vehicle image is determined according to the processed feature information, so as to improve the detection accuracy of the vehicle damage and the detection speed of the vehicle damage. For example, it can be used for auto insurance business in the financial field. For auto insurance business, when it is necessary to assess the damage of the vehicle, the vehicle damage information of the vehicle can be determined by the vehicle damage detection method. Accordingly, business personnel can promote auto insurance business based on the vehicle damage information of the vehicle, which is conducive to improving the promotion efficiency of the auto insurance business. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0028] Figure 1 It is a flow chart of a vehicle damage detection method provided in an embodiment of the present application;

[0029] Figure 2 is a schematic diagram of a vehicle damage detection method involved in an embodiment of the present application;

[0030] Figure 3 is a schematic diagram of the network architecture of a backbone network involved in an embodiment of the present application;

[0031] Figure 4 It is a flowchart of a method for training a vehicle damage detection model provided in an embodiment of the present application;

[0032] Figure 5 is a schematic block diagram of a vehicle damage detection device provided in an embodiment of the present application;

[0033] Figure 6 It is a schematic block diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0035] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0036] The embodiments of the present application provide a vehicle damage detection method, training method, device, equipment and storage medium. Among them, the vehicle damage detection method and the training method of the vehicle damage detection model can be applied to a computer device, which can be a mobile phone, a tablet computer, a laptop computer, a desktop computer and other devices. It can also be applied to a server, which can be a separate server or 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, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0037] In conjunction with the accompanying drawings, some embodiments of the present application are described in detail below. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0038] See also Figure 1 , Figure 1 It is a flow chart of a vehicle damage detection method provided in an embodiment of the present application. It should be noted that the vehicle damage detection method provided in an embodiment of the present application can be used in a computer device, and of course can also be used in a server. For example, the server can obtain a vehicle image from a computer device, process the vehicle image according to the vehicle damage detection method, and obtain vehicle damage information corresponding to the vehicle image. Exemplarily, the server can, for example, send the vehicle damage information to a computer device; of course, it is not limited to this and is not limited here.

[0039] In specific implementation, computer devices include but are not limited to: any one of mobile phones, tablet computers, laptop computers, and desktop computers; the server can be a single server or a server cluster, or a cloud server that provides cloud computing services.

[0040] like Figure 1 As shown, the vehicle damage detection method includes steps S101 to S104.

[0041] Step S101: Acquire a vehicle image.

[0042] When it is necessary to detect the vehicle damage information, the computer device can acquire the vehicle image.

[0043] For example, in the case of auto insurance, when a car owner applies for auto insurance claim, he or she can take a photo of the car to obtain a vehicle image. Of course, this is not limited to this. When a business person handles a car insurance claim, he or she can also take a photo of the car owner's car to obtain a vehicle image, which is not limited here.

[0044] In some embodiments, the computer device may output vehicle photography prompt information to prompt at least one of the vehicle owner and the business personnel to photograph the damaged area of ​​the vehicle. The vehicle photography prompt information may include text prompts, picture prompts, video prompts, animation prompts, etc., which are not limited here.

[0045] In some embodiments, when the computer device acquires the vehicle image, it may perform image preprocessing on the vehicle image to determine the image regions corresponding to different parts of the vehicle. For example, the vehicle image may be divided into an image region corresponding to the front of the vehicle, an image region corresponding to the body of the vehicle, an image region corresponding to the rear of the vehicle, etc., so as to subsequently determine the vehicle damage information corresponding to each of the different image regions, which is not limited here.

[0046] In this way, when a vehicle image is acquired, the vehicle image can be used to subsequently determine the vehicle damage information.

[0047] Step S102: Based on the backbone network of the vehicle damage detection model, feature extraction processing is performed on the vehicle image to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depthwise separable convolutional layer; the size of the convolution kernel in the depthwise separable convolutional layer is greater than or equal to a preset size threshold.

[0048] Exemplarily, the vehicle damage detection model may be designed based on a single-stage target detector, and the vehicle damage detection model may include a backbone network, a neck network, and a head network.

[0049] The computer device can input the acquired vehicle image into the vehicle damage detection model. The vehicle damage detection model can perform feature extraction processing on the vehicle image through the backbone network to obtain feature information corresponding to the vehicle image.

[0050] In some embodiments, when the backbone network performs feature extraction processing on the vehicle image, multiple layers of feature information of different scales can be extracted from the vehicle image. The information or structure represented by the feature information of different scales has different spatial sizes or granularities in the vehicle damage image. For example, the feature information of different scales describes the image content at different levels of detail or abstraction to capture different levels of information in the image. Among them, the multiple layers of feature information of different scales can be represented as a multi-scale feature pyramid, which is not limited here.

[0051] like Figure 2 As shown, when the backbone network receives a vehicle image, it can extract features of the vehicle image to obtain features C3, C4, and C5. Features C3, C4, and C5 can have different scales. Accordingly, features C3, C4, and C5 can be embodied as a multi-scale feature pyramid.

[0052] Exemplarily, the network architecture of the backbone network may include a depthwise separable convolution layer, and the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to a preset size threshold. For example, the network architecture of the backbone network may adopt a residual neural network architecture (Residual Neural Network, ResNet), which is not limited here.

[0053] In some embodiments, the size of the convolution kernel in the depthwise separable convolution layer may include 5×5, 7×7, etc. The size of the convolution kernel in the depthwise separable convolution layer may be preset or set by the user, and is not limited here. The preset size threshold may include 5×5, and the preset size threshold may be preset or set by the user, and is not limited here.

[0054] like Figure 3As shown, the backbone network can be provided with a depth-separable convolution layer with a convolution kernel size of 5×5. Of course, it is not limited to this. The backbone network can be provided with a convolution layer with a convolution kernel size of 3×3, a batch normalization (Batch Normalization, BN) layer corresponding to the 3×3 convolution layer, and a nonlinear activation function (Sigmoid Linear Unit, SiLU) before the depth-separable convolution layer. Correspondingly, the 5×5 depth-separable convolution layer can also be provided with a corresponding BN layer and SiLU function, which is not limited here. After the vehicle image is input into the backbone network of the vehicle damage detection model, it can be convolved by a 3×3 convolution layer, processed by the BN layer corresponding to the 3×3 convolution layer and activated by the SiLU function, and then input into a 5×5 depth-separable convolution layer for depth-separable convolution, and after being processed by the 5×5 depth-separable convolution layer BN layer and activated by the SiLU function, the feature information corresponding to the vehicle image is obtained.

[0055] The feature information corresponding to the vehicle image can be used by the vehicle damage detection model to subsequently determine the vehicle damage information corresponding to the vehicle image.

[0056] In this way, the vehicle damage detection model sets the network architecture of the backbone network to include a depthwise separable convolution layer, and the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to the preset size threshold, so that the backbone network can obtain a larger receptive field based on the depthwise separable convolution layer in the process of extracting feature information corresponding to the vehicle image, which helps the backbone network to more comprehensively capture and model the contextual information of the vehicle image, thereby facilitating the backbone network to extract feature information corresponding to the vehicle image. Accordingly, based on the improvement of the efficiency of extracting feature information corresponding to the vehicle image based on the backbone network, it is conducive to subsequently improving the accuracy of the vehicle damage detection model in detecting vehicle damage information corresponding to the vehicle image, and at the same time, it is conducive to improving the detection speed of vehicle damage information.

[0057] Step S103: Based on the neck network of the vehicle damage detection model, feature enhancement and fusion processing are performed on the feature information to obtain processed feature information; the network architecture of the neck network is the same as that of the backbone network.

[0058] After the vehicle damage detection model determines the feature information corresponding to the vehicle image through the backbone network, the acquired feature information can be input into the neck network, and the neck network can perform feature enhancement and fusion processing on the feature information to obtain the processed feature information.

[0059] In some embodiments, when the neck network receives feature information corresponding to the vehicle image output by the backbone network, the neck network can perform bottom-up and top-down feature propagation on the feature information to enhance the feature information, and perform fusion processing on the feature information to obtain processed feature information. Figure 2 As shown, when the neck network receives the features C3, C4 and C5 output by the backbone network, the neck network can perform corresponding concatenation (Concat) operations, resize (Resize) operations and convolution (Conv) operations on the features C3, C4 and C5, respectively, so as to enhance and fuse the features C3, C4 and C5, thereby outputting the processed features C3, C4 and C5, that is, outputting the processed feature information.

[0060] Exemplarily, the network architecture of the neck network is the same as the network architecture of the backbone network.

[0061] For example, the network architecture of the neck network and the network architecture of the backbone network both include a depthwise separable convolutional layer, and the size of the convolution kernel in the depthwise separable convolutional layer is greater than or equal to a preset size threshold.

[0062] like Figure 3 As shown, when the backbone network is provided with a depth-separable convolution layer with a convolution kernel size of 5×5, the neck network is also provided with a depth-separable convolution layer with a convolution kernel size of 5×5. Of course, it is not limited to this. When the backbone network is provided with a convolution layer with a convolution kernel size of 3×3, a BN layer corresponding to the 3×3 convolution layer, and a SiLU function, the neck network is also provided with a convolution layer with a convolution kernel size of 3×3, a BN layer corresponding to the 3×3 convolution layer, and a SiLU function. Correspondingly, when the backbone network is provided with a BN layer corresponding to a 5×5 depth-separable convolution layer and a SiLU function, the neck network is also provided with a BN layer corresponding to a 5×5 depth-separable convolution layer and a SiLU function, which is not limited here.

[0063] The network architecture based on the neck network is the same as the network architecture of the backbone network. The vehicle damage detection model can enhance the connectivity between the neck network and the backbone network, which is beneficial to improve the expansion rate of the basic blocks involved in the network architecture in the neck network. The basic blocks include, for example, at least one of a depth-separable convolutional layer, a convolutional layer, a BN layer, and a SiLU function, which are not limited here. Accordingly, based on the improvement in the expansion rate, the vehicle damage detection model can transfer some parameters and calculations related to the feature information from the backbone network to the neck network, and the neck network and the backbone network can have similar capacity, which is beneficial to balance the computational complexity of the feature information and improve the accuracy of determining the feature information.

[0064] In some embodiments, the network architecture of the backbone network and the network architecture of the neck network both include point-by-point convolution layers, and the number of basic blocks corresponding to the network architecture of the backbone network and the number of basic blocks corresponding to the network architecture of the neck network are both less than or equal to a preset number threshold; wherein the size of the convolution kernel in the point-by-point convolution layer is smaller than the size of the convolution kernel in the depthwise separable convolution layer.

[0065] For example, when the network architecture of the neck network and the network architecture of the backbone network both include a depthwise separable convolution layer, and the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to a preset size threshold, the vehicle damage detection model can set a pointwise convolution layer after the depthwise separable convolution layer, and the network architecture of the backbone network and the network architecture of the neck network both include a pointwise convolution layer. Among them, the size of the convolution kernel in the pointwise convolution layer is smaller than the size of the convolution kernel in the depthwise separable convolution layer.

[0066] like Figure 3 As shown, when the network architecture of the backbone network and the network architecture of the neck network both include a depthwise separable convolution layer with a convolution kernel size of 5×5, a pointwise convolution layer with a convolution kernel size of 1×1 can be set after the 5×5 depthwise separable convolution layer. Of course, it is not limited to this. The vehicle damage detection model can also set a corresponding BN layer and SiLU function after the 1×1 pointwise convolution layer, which is not limited here.

[0067] When the network architecture of the backbone network and the network architecture of the neck network both include point-by-point convolutional layers, the number of basic blocks corresponding to the network architecture of the backbone network and the number of basic blocks corresponding to the network architecture of the neck network will increase. As the number of basic blocks increases, the reasoning speed of the backbone network and the neck network for the corresponding feature information will decrease. Based on this, in order to ensure the reasoning speed of the backbone network and the neck network for the corresponding feature information, the vehicle damage detection model can limit the number of basic blocks corresponding to the network architecture of the backbone network and the number of basic blocks corresponding to the network architecture of the neck network.

[0068] For example, the vehicle damage detection model can set a preset quantity threshold corresponding to the number of basic blocks based on the impact of the number of basic blocks on the reasoning speed of the backbone network and the neck network for the corresponding feature information. The preset quantity threshold can be pre-set or can be adjusted and set by the user, which is not limited here. Based on the setting of the preset quantity threshold, the vehicle damage detection model can limit the number of basic blocks corresponding to the network architecture of the backbone network and the number of basic blocks corresponding to the network architecture of the neck network to be less than or equal to the preset quantity threshold.

[0069] In this way, the vehicle damage detection model can, while limiting the number of basic blocks corresponding to the network architecture of the backbone network and the number of basic blocks corresponding to the network architecture of the neck network, appropriately increase the convolution kernel size of the depth-separable convolution layer corresponding to the backbone network and the neck network, so that the backbone network and the neck network can parallelize the reasoning of the corresponding feature information, and can maintain the model capacity of the vehicle damage detection model. Based on this, the vehicle damage detection model can ensure the accuracy of the backbone network in extracting feature information, and the efficiency of the feature enhancement and fusion processing of the neck network on the feature information can be improved while ensuring the accuracy of the neck network in processing feature information, which is conducive to the subsequent improvement of the vehicle damage detection model's detection accuracy of the vehicle damage information corresponding to the vehicle image, and the improvement of the detection speed of the vehicle damage information.

[0070] Step S104: Based on the head network of the vehicle damage detection model, the vehicle damage information corresponding to the vehicle image is determined according to the processed feature information.

[0071] After the vehicle damage detection model determines the processed feature information through the neck network, the processed feature information can be input into the head network, and the head network can perform corresponding detection processing on the processed feature information to obtain the vehicle damage information corresponding to the vehicle image.

[0072] For example, the head network can perform border detection and instance segmentation on the processed feature information to determine the vehicle damage information corresponding to the vehicle image. Figure 2 As shown, when the head network receives the processed features C3, C4 and C5 output by the neck network, the head network can perform border detection and instance segmentation on the processed features C3, C4 and C5 to determine the vehicle damage information corresponding to the vehicle image. In the process of performing border detection on the processed features C3, C4 and C5, the head network can perform classification (Cls.) and regression (Reg.) processing on the corresponding feature information. In the process of performing instance segmentation on the processed features C3, C4 and C5, the head network can perform dynamic convolution processing (Daynamic Convs) on the corresponding feature information based on the self-supervised learning algorithm (Mask Feat).

[0073] In some embodiments, a first detection head based on the head network performs border detection on the processed feature information to obtain border information corresponding to the vehicle image; a second detection head based on the head network performs instance segmentation on the processed feature information to obtain mask information corresponding to the vehicle image; and based on the border information and the mask information, the vehicle damage information corresponding to the vehicle image is determined.

[0074] For example, the head network may be provided with a first detection head and a second detection head. The head network may perform border detection processing on the processed feature information through the first detection head to determine the border information corresponding to the vehicle image. Correspondingly, the head network may perform instance segmentation processing on the processed feature information through the second detection head to determine the mask information corresponding to the vehicle image. When the border information is determined, the vehicle damage detection model may frame the area where the vehicle damage information is located on the vehicle image. When the mask information is determined, the vehicle damage detection model may highlight the location where the vehicle damage information is located on the vehicle image. Based on this, the vehicle damage detection model may determine the vehicle damage information corresponding to the vehicle image according to the border information and the mask information. The vehicle damage detection model may output the vehicle image carrying the border information and the mask information when the vehicle damage information corresponding to the vehicle image is determined. For the vehicle insurance business, relevant personnel, such as business personnel, for example, may view the vehicle damage information corresponding to the vehicle image according to the vehicle image carrying the border information and the mask information, so as to carry out damage assessment processing for the vehicle corresponding to the vehicle image, which is not limited here.

[0075] In this way, through the setting of the first detection head and the second detection head, the head network can detect the border information corresponding to the vehicle image and the mask information corresponding to the vehicle image in parallel to determine the vehicle damage information corresponding to the vehicle image, which is beneficial to improve the detection accuracy of the vehicle damage detection model for vehicle damage information while improving the detection speed of vehicle damage information.

[0076] Exemplarily, the feature information includes multiple feature information of different scales. Correspondingly, the processed feature information may also include multiple processed feature information of different scales. Figure 2 As shown, the feature information may include feature C3, feature C4 and feature C5, and feature C3, feature C4 and feature C5 have different scales, and the processed feature C3, the processed feature C4 and the processed feature C5 also have different scales.

[0077] In some embodiments, the processed multiple feature information of different scales are input into the first detection head, so that the first detection head performs border detection processing on the processed multiple feature information of different scales according to preset head parameters to obtain border information.

[0078] After the vehicle damage detection model determines the processed multiple feature information of different scales through the neck network, the processed multiple feature information of different scales can be input into the first detection head, and the first detection head can perform border detection processing on the processed multiple feature information of different scales according to the preset head parameters to obtain border information. Since the first detection head needs to process the processed multiple feature information of different scales according to the same preset head parameters, the preset head parameters can also be called cross-scale shared head parameters. Correspondingly, based on the cross-scale shared head parameters adopted by the first detection head, the first detection head can use the BN layer with shared convolution weights and separation to predict the classification and regression results of the border information to obtain the border information. The preset head parameters can be pre-set or set by the user. For example, the user can determine the parameters of the BN layer corresponding to the first detection head according to the statistical data of the vehicle damage detection model during the training process, thereby determining the preset head parameters of the first detection head, which is not limited here.

[0079] In this way, based on the setting of the preset head parameters, the parameter amount of the first detection head can be reduced, which is conducive to reducing the parameter amount of the head network of the vehicle damage detection model. The reduction of the parameter amount of the head network is conducive to improving the detection speed of the vehicle damage detection model for vehicle damage information.

[0080] The vehicle damage detection method provided in the above embodiment obtains a vehicle image; based on the backbone network of the vehicle damage detection model, the vehicle image is subjected to feature extraction processing to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depth-separable convolution layer; the size of the convolution kernel in the depth-separable convolution layer is greater than or equal to the preset size threshold; based on the neck network of the vehicle damage detection model, the feature information is subjected to feature enhancement and fusion processing to obtain processed feature information; the network architecture of the neck network is the same as the network architecture of the backbone network; based on the head network of the vehicle damage detection model, the vehicle damage information corresponding to the vehicle image is determined according to the processed feature information, so as to improve the detection accuracy of the vehicle damage and the detection speed of the vehicle damage. For example, it can be used in the auto insurance business in the financial field. For the auto insurance business, when it is necessary to assess the damage to the vehicle, the vehicle damage information of the vehicle can be determined by the vehicle damage detection method. Accordingly, the business personnel can promote the auto insurance business based on the vehicle damage information, which is conducive to improving the promotion efficiency of the auto insurance business.

[0081] See also Figure 4 , Figure 4It is a flow chart of a method for training a vehicle damage detection model provided in an embodiment of the present application. It should be noted that the method for training a vehicle damage detection model provided in an embodiment of the present application can be used in a computer device, and of course can also be used in a server. For example, the server can obtain a training sample set from a computer device, process the training sample set according to the training method of the vehicle damage detection model, and obtain model optimization parameters for the vehicle damage detection model. And according to the model optimization parameters, adjust the model parameters of the vehicle damage detection model to obtain an optimized vehicle damage detection model. Exemplarily, the server can, for example, send the optimized vehicle damage detection model to a computer device, so that when the computer device obtains a vehicle image, it can determine the vehicle damage information corresponding to the vehicle image through the optimized vehicle damage detection model; of course, it is not limited to this, and no limitation is made here.

[0082] In specific implementation, computer devices include but are not limited to: any one of mobile phones, tablet computers, laptop computers, and desktop computers; the server can be a single server or a server cluster, or a cloud server that provides cloud computing services.

[0083] like Figure 4 As shown, the training method of the vehicle damage detection model includes steps S201 to S206.

[0084] Step S201: Obtain a training sample set, where the training sample set includes a plurality of vehicle images and vehicle damage information labels corresponding to the plurality of vehicle images.

[0085] For example, when training a vehicle damage detection model, a training sample set may be obtained. The training sample set may include multiple vehicle images. Each vehicle image in the training sample set may be provided with a corresponding vehicle damage information label. The training sample set may be used for subsequent training of the vehicle damage detection model.

[0086] In some embodiments, when a vehicle image is acquired, the vehicle damage information label corresponding to the vehicle image can be determined by a preset annotation method, which is not limited here.

[0087] In this way, when the vehicle image and the vehicle damage information label corresponding to the vehicle image are obtained, it can be used for subsequent training of the vehicle damage detection model.

[0088] Step S202: Based on the backbone network of the vehicle damage detection model, feature extraction processing is performed on the vehicle image to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depthwise separable convolutional layer; the size of the convolution kernel in the depthwise separable convolutional layer is greater than or equal to a preset size threshold.

[0089] For example, when training a vehicle damage detection model, it is necessary to learn how to determine the feature information corresponding to a vehicle image. In some embodiments, the vehicle damage detection model may be provided with a backbone network, the network architecture of the backbone network includes a depthwise separable convolution layer, and the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to a preset size threshold, and the backbone network may be used to extract and process the feature information corresponding to the vehicle image. Based on this, the vehicle image may be input into the backbone network of the vehicle damage detection model, so that the backbone network may perform feature extraction processing on the vehicle image to obtain the feature information corresponding to the vehicle image.

[0090] In this way, when the feature information corresponding to the vehicle image is determined, the vehicle damage detection model can optimize the model parameters of the vehicle damage detection model according to the feature information corresponding to the vehicle image to improve the training effect of the vehicle damage detection model.

[0091] Step S203: Based on the neck network of the vehicle damage detection model, feature enhancement and fusion processing are performed on the feature information to obtain processed feature information; the network architecture of the neck network is the same as the network architecture of the backbone network.

[0092] For example, when training a vehicle damage detection model, it is necessary to learn how to perform feature enhancement and fusion processing on feature information. In some embodiments, the vehicle damage detection model may be provided with a neck network, and the network architecture of the neck network is the same as the network architecture of the backbone network. When the network architecture of the backbone network includes a depthwise separable convolution layer, and the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to a preset size threshold, the network architecture of the neck network also includes a depthwise separable convolution layer, and the size of the convolution kernel in the depthwise separable convolution layer is also greater than or equal to a preset size threshold. Of course, the network architecture of the backbone network and the network architecture of the neck network are not limited to this, and are not limited here. Based on this, the feature information corresponding to the vehicle image can be input into the neck network of the vehicle damage detection model, so that the neck network can perform feature enhancement and fusion processing on the feature information to obtain the processed feature information.

[0093] In this way, when the processed feature information corresponding to the vehicle image is determined, the vehicle damage detection model can optimize the model parameters of the vehicle damage detection model based on the processed feature information to improve the training effect of the vehicle damage detection model.

[0094] Step S204: Based on the head network of the vehicle damage detection model, the vehicle damage information corresponding to the vehicle image is determined according to the processed feature information.

[0095] For example, when training a vehicle damage detection model, it is necessary to learn how to determine the vehicle damage information corresponding to a vehicle image. In some embodiments, the vehicle damage detection model may be provided with a head network. For example, it is necessary to learn how to determine the vehicle damage information corresponding to a vehicle image based on processed feature information. Based on this, the processed feature information may be input into the head network of the vehicle damage detection model so that the head network can determine the vehicle damage information corresponding to the vehicle image.

[0096] In some embodiments, the head network may include a first detection head and a second detection head. When training the vehicle damage detection model, the first detection head may perform border detection on the processed feature information to learn how to determine the border information corresponding to the vehicle image. Accordingly, when training the vehicle damage detection model, the second detection head may perform instance segmentation on the processed feature information to learn how to determine the mask information corresponding to the vehicle image. When determining the border information corresponding to the vehicle information and the mask information corresponding to the vehicle information, the head network may learn how to determine the vehicle damage information corresponding to the vehicle image based on the border information and the mask information.

[0097] In this way, when the vehicle damage information corresponding to the vehicle image is determined, the vehicle damage detection model can optimize the model parameters of the vehicle damage detection model based on the vehicle damage information to improve the training effect of the vehicle damage detection model.

[0098] Step S205: Determine the model optimization parameters according to the intersection-and-union ratio between the vehicle damage information and the vehicle damage information label.

[0099] For example, in the process of training the vehicle damage detection model, a dynamic label allocation strategy, such as a training strategy of softened label allocation, can be used to train the vehicle damage detection model. For example, the vehicle damage detection model can use the intersection over union (IOU) between the vehicle damage information and the vehicle damage information label as a soft label to determine the model optimization parameters corresponding to the vehicle damage detection model. Among them, the IOU between the vehicle damage information and the vehicle damage information can be expressed as Y soft .

[0100] For example, the vehicle damage information corresponding to the vehicle image determined by the vehicle damage detection model is equivalent to the predicted result, and the vehicle damage information label corresponding to the vehicle image included in the sample data set is equivalent to the actual result. The vehicle damage detection model can determine the intersection-and-union ratio between the vehicle damage information and the vehicle damage information label based on the intersection-and-union ratio between the bounding box information indicated by the vehicle damage information and the bounding box information indicated by the vehicle damage information label, so as to determine the model optimization parameters.

[0101] In some implementations, a cost function corresponding to the vehicle damage detection model is determined based on the intersection-to-union ratio; based on the cost function, the vehicle damage detection model is iteratively optimized to obtain model optimization parameters corresponding to the vehicle damage detection model.

[0102] For example, when the intersection-over-union ratio is obtained, the computer device can determine the cost function corresponding to the vehicle damage detection model. The cost function can be used to calculate the cost between the vehicle damage information determined by the vehicle damage detection model and the soft label, so as to provide for subsequent iterative optimization of the vehicle damage detection model.

[0103] Exemplarily, the classification cost corresponding to the vehicle damage detection model is determined according to the intersection-and-union ratio; the regression cost corresponding to the vehicle damage detection model is determined according to the logarithm of the intersection-and-union ratio; the regional cost corresponding to the vehicle damage detection model is determined according to the soft center area cost corresponding to the intersection-and-union ratio; the cost function is determined according to the classification cost, regression cost, regional cost and the weight coefficients corresponding to the classification cost, regression cost and regional cost.

[0104] For example, the classification cost can be expressed as C cls , C cls As shown below:

[0105] C cls =CE(P,Y soft) ×(Y soft -P) 2

[0106] Among them, P is used to indicate the probability distribution information of vehicle damage information determined by the vehicle damage detection model, and Y soft It is used to indicate the intersection-over-union ratio between vehicle damage information and vehicle damage information labels, and CE is used to indicate the cross entropy loss function.

[0107] When soft labels are introduced into the cost function, the classification costs of different regression qualities of the vehicle damage detection model can be re-weighted, which is helpful to avoid noisy and unstable matching caused by binary labels, thereby facilitating the subsequent improvement of the training effect of the vehicle damage detection model.

[0108] For example, the regression cost can be expressed as C reg , C reg As shown below:

[0109] C reg = -log(IOU)

[0110] Since the regression cost is determined based on the logarithm of the IOU, when the commonly used GIOU is replaced by the logarithm of the IOU, the cost of the classification with a lower IOU value can be amplified during the training of the vehicle damage detection model, which is conducive to the subsequent improvement of the training effect of the vehicle damage detection model.

[0111] For example, the regional cost can be expressed as C center , C center As shown below:

[0112]

[0113] Among them, δ and ε are hyperparameters indicating the cost of the soft center region, x pred Used to indicate the vehicle damage information corresponding to the vehicle image by the vehicle damage detection model, x gt Used to indicate vehicle damage information label.

[0114] In an exemplary implementation, δ may be set to 10, and ε may be set to 3, but this is certainly not limited thereto and is not limited here.

[0115] Since the regional cost is determined according to the soft center regional cost corresponding to the intersection-union ratio, when the vehicle damage detection model is trained by using the soft center regional cost instead of the fixed prior center, it is beneficial to improve the stability of the dynamic allocation cost of the soft label, which is beneficial to the subsequent improvement of the training effect of the vehicle damage detection model.

[0116] In determining the classification cost C cls , regression cost C reg and the regional cost C center In this case, computer equipment can be combined with the classification cost C cls , regression cost C reg and the regional cost C center The corresponding weight coefficients determine the cost function. The cost function can be expressed as function C, which is shown as follows:

[0117] C=α1C cls+ α2C reg +α3C center

[0118] Among them, α1 is used to represent the weight coefficient corresponding to the classification cost, α2 is used to represent the weight coefficient corresponding to the regression cost, and α3 is used to represent the weight coefficient corresponding to the regional cost.

[0119] In an exemplary implementation, α1 may be set to 1, α2 may be set to 3, and α3 may be set to 1, but this is not limiting and is not limited here.

[0120] When the cost function is determined, the computer device can determine the model optimization parameters corresponding to the vehicle damage detection model according to the cost function. The model optimization parameters can be used to subsequently adjust the model parameters of the vehicle damage detection model so that the vehicle damage detection model can more accurately bring the prediction results of the vehicle damage information corresponding to the vehicle image of the vehicle damage detection model closer to the actual results, thereby achieving the purpose of training.

[0121] Step S206: adjusting the model parameters of the vehicle damage detection model according to the model optimization parameters.

[0122] For example, the vehicle damage detection model can adjust the model parameters of at least one of the backbone network, neck network, and head network of the vehicle damage detection model according to the model optimization parameters, which is not limited here.

[0123] Through the model parameters of the vehicle damage detection model, the vehicle damage detection model can more accurately bring the predicted results of the vehicle damage information corresponding to the vehicle image closer to the actual results, so as to achieve the purpose of training, which is conducive to improving the accuracy of the vehicle damage detection model in determining the vehicle damage information corresponding to the vehicle image.

[0124] The training method of the vehicle damage detection model provided in the above embodiment obtains a training sample set, the training sample set includes multiple vehicle images and vehicle damage information labels corresponding to each of the multiple vehicle images; based on the backbone network of the vehicle damage detection model, the vehicle image is subjected to feature extraction processing to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depth-separable convolution layer; the size of the convolution kernel in the depth-separable convolution layer is greater than or equal to a preset size threshold; based on the neck network of the vehicle damage detection model, the feature information is subjected to feature enhancement and fusion processing to obtain processed feature information; the network architecture of the neck network is the same as the network architecture of the backbone network; based on the head network of the vehicle damage detection model, the vehicle damage information corresponding to the vehicle image is determined according to the processed feature information; according to the intersection-and-union ratio between the vehicle damage information and the vehicle damage information label, the model optimization parameters are determined, and according to the model optimization parameters, the model parameters of the vehicle damage detection model are adjusted to improve the accuracy of determining the vehicle damage information corresponding to the vehicle image, thereby improving the training effect of the vehicle damage detection model. The vehicle damage detection model can be subsequently applied to the vehicle image, which is conducive to improving the accuracy of determining the vehicle damage information corresponding to the vehicle image. For example, it can be used in the auto insurance business in the financial field. For vehicle images corresponding to vehicles that need to be assessed, the vehicle damage information corresponding to the vehicle images can be detected and determined through the trained vehicle damage detection model, which is conducive to improving the convenience of vehicle damage assessment. In addition, it is not necessary to spend a lot of manpower and time to determine the vehicle damage information, which is conducive to improving the efficiency of promoting auto insurance business.

[0125] See also Figure 5 , Figure 5 1 is a schematic block diagram of a vehicle damage detection device provided in an embodiment of the present application. The vehicle damage detection device can be configured in a server or an electronic device to execute the aforementioned vehicle damage detection method.

[0126] like Figure 5 As shown, the vehicle damage detection device includes: an image acquisition module 110, a feature extraction module 120, a feature processing module 130 and a vehicle damage identification module 140.

[0127] The image acquisition module 110 is used to acquire vehicle images.

[0128] The feature extraction module 120 is used to perform feature extraction processing on the vehicle image based on the backbone network of the vehicle damage detection model to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depthwise separable convolutional layer; the size of the convolution kernel in the depthwise separable convolutional layer is greater than or equal to a preset size threshold.

[0129] The feature processing module 130 is used to perform feature enhancement and fusion processing on the feature information based on the neck network of the vehicle damage detection model to obtain processed feature information; the network architecture of the neck network is the same as the network architecture of the backbone network.

[0130] The vehicle damage recognition module 140 is used to determine the vehicle damage information corresponding to the vehicle image based on the head network of the vehicle damage detection model according to the processed feature information.

[0131] It should be noted that technicians in the relevant field can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described devices and modules and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0132] The method of the present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0133] Exemplarily, the above method and apparatus may be implemented in the form of a computer program, which may be run on a computer device.

[0134] See also Figure 6 , Figure 6 1 is a schematic block diagram of a computer device provided in an embodiment of the present application. The computer device may be a server or an electronic device.

[0135] like Figure 6 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a storage medium and an internal memory.

[0136] The storage medium can store an operating system and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any vehicle damage detection method, or can execute any vehicle damage detection model training method.

[0137] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0138] The internal memory provides an environment for the operation of the computer program in the storage medium. When the computer program is executed by the processor, the processor can execute any vehicle damage detection method, or can execute any vehicle damage detection model training method.

[0139] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0140] It should be understood that the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0141] In one embodiment, the processor is used to execute a computer program and implement the following steps when executing the computer program:

[0142] Acquire vehicle images;

[0143] Based on the backbone network of the vehicle damage detection model, feature extraction processing is performed on the vehicle image to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depthwise separable convolution layer; the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to a preset size threshold;

[0144] Based on the neck network of the vehicle damage detection model, feature enhancement and fusion processing are performed on the feature information to obtain processed feature information; the network architecture of the neck network is the same as the network architecture of the backbone network;

[0145] Based on the head network of the vehicle damage detection model, the vehicle damage information corresponding to the vehicle image is determined according to the processed feature information.

[0146] It should be noted that technicians in the relevant field can clearly understand that for the convenience and brevity of description, the specific working process of vehicle damage detection described above can refer to the corresponding process in the aforementioned vehicle damage detection method embodiment, and will not be repeated here.

[0147] In another embodiment, the processor is used to execute a computer program and implement the following steps when executing the computer program:

[0148] Acquire a training sample set, wherein the training sample set includes a plurality of vehicle images and vehicle damage information labels corresponding to the plurality of vehicle images;

[0149] Based on the backbone network of the vehicle damage detection model, feature extraction processing is performed on the vehicle image to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depthwise separable convolution layer; the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to a preset size threshold;

[0150] Based on the neck network of the vehicle damage detection model, feature enhancement and fusion processing are performed on the feature information to obtain processed feature information; the network architecture of the neck network is the same as the network architecture of the backbone network;

[0151] Based on the head network of the vehicle damage detection model, determining the vehicle damage information corresponding to the vehicle image according to the processed feature information;

[0152] Determining a model optimization parameter according to an intersection-and-union ratio between the vehicle damage information and the vehicle damage information label;

[0153] According to the model optimization parameters, the model parameters of the vehicle damage detection model are adjusted.

[0154] It should be noted that technical personnel in the relevant field can clearly understand that, for the convenience and conciseness of description, the specific working process of the training of the vehicle damage detection model described above can refer to the corresponding process in the aforementioned vehicle damage detection model training method embodiment, and will not be repeated here.

[0155] An embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. The method implemented when the computer program is executed by a processor can refer to the various embodiments of the vehicle damage detection method of the present application, or can refer to the various embodiments of the training method of the vehicle damage detection model of the present application.

[0156] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiment, such as a hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, a smart memory card (SmartMedia Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc., equipped on the computer device.

[0157] It should be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0158] It should also be understood that the term "and / or" used in the specification of this application and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, including these combinations. It should be noted that, in this article, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "including a..." does not exclude the presence of other identical elements in the process, method, article or system including the element.

[0159] The serial numbers of the embodiments of the present application are for description only and do not represent the advantages and disadvantages of the embodiments. The above description is only a specific implementation mode of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present application, and these modifications or replacements should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A vehicle damage detection method, characterized in that: include: Acquire vehicle images; Based on the backbone network of the vehicle damage detection model, feature extraction processing is performed on the vehicle image to obtain feature information corresponding to the vehicle image; The network architecture of the backbone network includes a depthwise separable convolutional layer; the size of the convolution kernel in the depthwise separable convolutional layer is greater than or equal to a preset size threshold; Based on the neck network of the vehicle damage detection model, feature enhancement and fusion processing are performed on the feature information to obtain processed feature information; The network architecture of the neck network is the same as the network architecture of the backbone network; Based on the head network of the vehicle damage detection model, the vehicle damage information corresponding to the vehicle image is determined according to the processed feature information.

2. The vehicle damage detection method according to claim 1, characterized in that: The network architecture of the backbone network and the network architecture of the neck network both include point-by-point convolutional layers, and the number of basic blocks corresponding to the network architecture of the backbone network and the number of basic blocks corresponding to the network architecture of the neck network are both less than or equal to a preset number threshold; wherein the size of the convolution kernel in the point-by-point convolutional layer is smaller than the size of the convolution kernel in the depth-wise separable convolutional layer.

3. The vehicle damage detection method according to claim 1 or 2, characterized in that: The head network based on the vehicle damage detection model determines the vehicle damage information corresponding to the vehicle image according to the processed feature information, including: Based on the first detection head of the head network, performing border detection processing on the processed feature information to obtain border information corresponding to the vehicle image; Based on the second detection head of the head network, performing instance segmentation processing on the processed feature information to obtain mask information corresponding to the vehicle image; Vehicle damage information corresponding to the vehicle image is determined according to the frame information and the mask information.

4. The vehicle damage detection method according to claim 3, characterized in that: The feature information includes multiple layers of feature information at different scales; The first detection head based on the head network performs border detection processing on the processed feature information to obtain border information corresponding to the vehicle image, including: The processed multi-layer feature information of different scales is input into the first detection head, so that the first detection head performs border detection processing on the processed multi-layer feature information of different scales according to preset head parameters to obtain the border information.

5. A method for training a vehicle damage detection model, characterized in that: include: Acquire a training sample set, wherein the training sample set includes a plurality of vehicle images and vehicle damage information labels corresponding to the plurality of vehicle images; Based on the backbone network of the vehicle damage detection model, feature extraction processing is performed on the vehicle image to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depthwise separable convolution layer; the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to a preset size threshold; Based on the neck network of the vehicle damage detection model, feature enhancement and fusion processing are performed on the feature information to obtain processed feature information; The network architecture of the neck network is the same as the network architecture of the backbone network; Based on the head network of the vehicle damage detection model, determining the vehicle damage information corresponding to the vehicle image according to the processed feature information; Determining a model optimization parameter according to an intersection-and-union ratio between the vehicle damage information and the vehicle damage information label; According to the model optimization parameters, the model parameters of the vehicle damage detection model are adjusted.

6. The training method according to claim 5, characterized in that: The determining of the model optimization parameters according to the intersection-and-union ratio between the vehicle damage information and the vehicle damage information label includes: Determining a cost function corresponding to the vehicle damage detection model according to the intersection-to-union ratio; Based on the cost function, the vehicle damage detection model is iteratively optimized to obtain model optimization parameters corresponding to the vehicle damage detection model.

7. The training method according to claim 6, characterized in that: Determining the cost function corresponding to the vehicle damage detection model according to the intersection-to-union ratio includes: Determining a classification cost corresponding to the vehicle damage detection model according to the intersection-and-union ratio; Determining a regression cost corresponding to the vehicle damage detection model according to the logarithm of the intersection-over-union ratio; Determining the regional cost corresponding to the vehicle damage detection model according to the soft center regional cost corresponding to the intersection-to-union ratio; The cost function is determined according to the classification cost, the regression cost, the regional cost and the weight coefficients corresponding to the classification cost, the regression cost and the regional cost respectively.

8. A vehicle damage detection device, characterized in that: The vehicle damage detection device comprises: An image acquisition module, used for acquiring vehicle images; A feature extraction module is used to perform feature extraction processing on the vehicle image based on the backbone network of the vehicle damage detection model to obtain feature information corresponding to the vehicle image; the network architecture of the backbone network includes a depthwise separable convolution layer; the size of the convolution kernel in the depthwise separable convolution layer is greater than or equal to a preset size threshold; A feature processing module, used for performing feature enhancement and fusion processing on the feature information based on the neck network of the vehicle damage detection model to obtain processed feature information; the network architecture of the neck network is the same as the network architecture of the backbone network; A vehicle damage recognition module is used to determine the vehicle damage information corresponding to the vehicle image based on the head network of the vehicle damage detection model and according to the processed feature information.

9. A computer device, characterized in that: The computer device includes a memory and a processor; The memory is used to store computer programs; The processor is used to execute the computer program and implement the vehicle damage detection method as described in any one of claims 1 to 4 when executing the computer program, or implement the vehicle damage detection model training method as described in any one of claims 5 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the vehicle damage detection method as described in any one of claims 1 to 4 are implemented, or the steps of the vehicle damage detection model training method as described in any one of claims 5 to 7 are implemented.

Images