Road quality testing methods, devices and related products

By improving the feature extraction of the YOLOv5 model and strengthening the network, the problem of low detection accuracy in pavement quality detection is solved, efficient detection of pavement diseases is achieved, and detection accuracy and feature extraction capabilities are improved.

CN115035305BActive Publication Date: 2024-11-15CAMBRICON SINGGO (NANJING) TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210700323.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-20
Publication Date
2024-11-15
Estimated Expiration
2042-06-20

AI Technical Summary

Technical Problem

The existing deep learning models failed to adapt to pavement disease characteristics in pavement quality detection, resulting in low detection accuracy.

Method used

The improved YOLOv5 model is adopted to detect pavement diseases through the improved feature extraction network and feature enhancement network. The feature extraction network has a receptive field size suitable for pavement disease detection, and the feature enhancement network integrates and strengthens the processing of medium and large-size features.

Benefits of technology

It improves the accuracy of pavement quality detection, is suitable for pavement disease detection of different types and sizes, and enhances the ability to extract medium and large-size disease characteristics.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115035305B_ABST
    Figure CN115035305B_ABST
Patent Text Reader

Abstract

The present application provides a road quality detection method, device and related products, which relate to the field of artificial intelligence technology. The specific scheme includes: obtaining a target image to be identified, wherein the target image includes a road surface area; using the improved feature extraction network in the YOLOv5 model to perform feature extraction processing on the target image to obtain extracted feature data; the improved feature extraction network has a receptive field size that is most suitable for road disease detection; using the improved feature enhancement network in the YOLOv5 model to perform feature enhancement processing on the feature data to obtain enhanced features; the enhanced features are the features after the feature data is initially fused and enhanced, and then the medium and large size features are fused and enhanced; the road surface diseases in the target image are detected according to the enhanced features. This scheme can effectively improve the accuracy of road quality detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a road quality detection method, device and related products. Background Art

[0002] With the continuous development of economy, traffic roads such as expressway network, urban road network, township road network have also been developed rapidly. Efficient and accurate detection of road pavement quality is the key to road maintenance.

[0003] At present, ordinary deep learning models are generally used for road quality detection. Ordinary deep learning models are not targeted at the scene of road quality detection. The model is adaptively adjusted based on the characteristics of road diseases, resulting in low accuracy in road quality detection. Summary of the invention

[0004] The embodiments of the present application provide a road quality detection method, device and related products to solve the technical problem that ordinary deep learning models in the prior art are not targeted at road quality detection scenarios, and the model is adaptively adjusted based on the characteristics of road surface diseases, resulting in low accuracy in road quality detection.

[0005] In a first aspect, the present application provides a road quality detection method, which acquires a target image to be identified, wherein the target image includes a road surface area;

[0006] Using an improved feature extraction network in a YOLOv5 model to perform feature extraction processing on the target image to obtain extracted feature data; the improved feature extraction network has a receptive field size suitable for pavement disease detection;

[0007] The feature data is subjected to feature enhancement processing by using the improved feature enhancement network in the YOLOv5 model to obtain enhanced features; the enhanced features are the features obtained by performing feature fusion enhancement processing on the medium and large size features after performing preliminary fusion enhancement processing on the feature data;

[0008] The road surface defects in the target image are detected according to the enhanced features.

[0009] In a second aspect, the present application provides a road surface quality detection device, comprising:

[0010] An acquisition module, used for acquiring a target image to be identified, wherein the target image includes a road surface area;

[0011] A feature extraction module, used for performing feature extraction processing on the target image using an improved feature extraction network in a YOLOv5 model to obtain extracted feature data; the improved feature extraction network has a receptive field size suitable for pavement disease detection;

[0012] A feature enhancement module is used to perform feature enhancement processing on the feature data using the improved feature enhancement network in the YOLOv5 model to obtain enhanced features; the enhanced features are the features obtained by performing feature fusion enhancement on the medium and large size features after performing preliminary fusion enhancement processing on the feature data;

[0013] A detection module is used to detect road surface defects in the target image based on the enhanced features.

[0014] In a third aspect, an embodiment of the present application provides a road quality detection device, comprising: a processor and a memory; wherein:

[0015] The memory is used to store program codes;

[0016] The processor is used to call the program code stored in the memory to execute the method described in the first aspect.

[0017] In a fourth aspect, an embodiment of the present application provides an artificial intelligence chip, which includes a road surface quality detection device as described in the second aspect or the third aspect.

[0018] In a fifth aspect, an embodiment of the present application provides a board card, which includes: a storage device, an interface device, a control device and an artificial intelligence chip as described in the fourth aspect.

[0019] In a sixth aspect, an embodiment of the present application provides an electronic device, comprising the artificial intelligence chip described in the fourth aspect.

[0020] In a seventh aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the computer executes the method described in the first aspect.

[0021] In an eighth aspect, an embodiment of the present application provides a computer program product, including a computer program, which implements the method described in the first aspect when executed by a processor.

[0022] The pavement quality detection method, device, and related products provided in the embodiments of the present application are obtained by acquiring a target image to be identified, wherein the target image includes a pavement area; using an improved feature extraction network in the YOLOv5 model to perform feature extraction processing on the target image to obtain extracted feature data; the improved feature extraction network has a receptive field size suitable for pavement disease detection; using an improved feature enhancement network in the YOLOv5 model to perform feature enhancement processing on the feature data to obtain enhanced features; the enhanced features are features obtained by performing feature fusion enhancement processing on medium and large-sized features after preliminary fusion enhancement processing of the feature data; pavement diseases in the target image are detected based on the enhanced features, and since the improved feature extraction network has a receptive field size suitable for pavement disease detection, it has an optimal target coverage rate suitable for pavement diseases of different types and sizes, and can effectively extract pavement disease features. Since the improved feature enhancement network performs preliminary fusion and enhancement processing on the feature data, and then performs feature fusion and enhancement on the medium and large-sized features, it can effectively enhance the characteristics of medium and large-sized pavement diseases. By improving the feature extraction network and feature enhancement network in the YOLOv5 model, the improved YOLOv5 model can be more suitable for pavement quality detection scenarios, thereby effectively improving the accuracy of pavement quality detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0024] Figure 1a-1d A schematic diagram of four types of pavement diseases in the pavement quality detection method provided in an embodiment of the present application;

[0025] Figure 2 A diagram of an application scenario of a road surface quality detection method provided by an embodiment of the present application;

[0026] Figure 3 A flow chart of a road surface quality detection method provided in one embodiment of the present application;

[0027] Figure 4 Schematic diagram of the YOLOv5 model;

[0028] Figure 5 A flow chart of a road surface quality detection method provided in yet another embodiment of the present application;

[0029] Figure 6 A flow chart of a road surface quality detection method provided in another embodiment of the present application;

[0030] Figure 7A flow chart of a road surface quality detection method provided in another embodiment of the present application;

[0031] Figure 8 A schematic diagram of an improved YOLOv5 model provided in an embodiment of the present application;

[0032] Fig. 9 A schematic diagram of the data processing process of the ECAM module provided in an embodiment of the present application;

[0033] Fig.10 A schematic diagram of the processing process of the ECA-CSP structure provided in an embodiment of the present application;

[0034] Fig.11 A flow chart of a road surface quality detection method provided in another embodiment of the present application;

[0035] Fig.12 A schematic diagram of the processing process of the improved feature enhancement network provided in an embodiment of the present application;

[0036] Fig.13 A schematic diagram of the structure of a road surface quality detection device provided in one embodiment of the present application;

[0037] Fig.14 A schematic diagram of the structure of a road surface quality detection device provided in another embodiment of the present application;

[0038] Fig.15 is a structural diagram showing a board according to an embodiment of the present application;

[0039] Fig.16 is a structural diagram showing a combined processing device according to an embodiment of the present application;

[0040] Fig.17 Schematic diagram showing the internal structure of a single-core computing device according to an embodiment of the present application

[0041] Fig.18 is a schematic diagram showing the internal structure of a multi-core computing device according to an embodiment of the present application;

[0042] Fig.19 Schematic diagram showing the internal structure of a processor core according to an embodiment of the present application.

[0043] The above drawings show clear embodiments of the present disclosure, which will be described in more detail below. These drawings and text descriptions are not intended to limit the scope of the present disclosure in any way, but to illustrate the concepts of the present disclosure to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0044] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0045] In order to clearly understand the technical solution of the present application, the solution of the prior art is first introduced in detail.

[0046] In the prior art, when detecting road quality, ordinary deep learning models are generally used. For example, it can be a deep learning model of a two-stage detection framework, and exemplary models can be: RCNN network, SPP network, Fast-RCNN network, etc. It can also be a deep learning model of a single-stage detection framework. Exemplarily, it can be: the first to fifth generation YOLO network model. However, these ordinary deep learning models are not targeted at the scenario of road quality detection, and the model is not adaptively adjusted based on the characteristics of road diseases, resulting in a low accuracy rate in detecting road quality.

[0047] Therefore, in the face of the problem that the common deep learning model in the existing technology has a low accuracy rate in detecting road quality, it is necessary to improve the existing common deep learning model. Since the fifth-generation YOLO model YOLOv5 has a good detection effect on target detection, the YOLOv5 model can be improved to make the improved YOLOv5 model more suitable for road quality detection. Specifically, by analyzing the characteristics of road diseases, such as Figure 1a to Figure 1d As shown in the figure, the main types of road surface diseases are vertical cracks, horizontal cracks, crocodile patterns and potholes. Among these road surface diseases, vertical cracks, horizontal cracks and crocodile patterns are all medium and large targets. The feature extraction network needs to have a sufficient receptive field to extract the feature information of the target. Therefore, it is necessary to improve the feature extraction network so that the improved feature extraction network has a receptive field size suitable for shoulder disease detection. In the YOLOv5 model, the feature enhancement network does not focus on distinguishing feature sizes, and comprehensively enhances the features. Since there are many medium and large targets in road surface diseases, it is necessary to improve the feature enhancement network so that the improved feature enhancement network performs preliminary fusion and enhancement processing on the feature data, and then performs feature fusion and enhancement on the medium and large size features. The features can not only fuse and enhance the small size features, but also effectively fuse and enhance the medium and large size features. By improving the feature extraction network and feature enhancement network in the YOLOv5 model, the improved YOLOv5 model can be more suitable for the scene of road quality detection, thereby effectively improving the accuracy of road quality detection.

[0048] The following is an explanation of the application scenarios of the road quality detection method provided in this application. Figure 2 As shown, in this embodiment, a video including a road surface area to be detected or a target image after decoding the video may be stored in the database 10. An improved YOLOv5 model is stored in the artificial intelligence chip 11, the board 12 or the electronic device 13. By inputting the target image into the artificial intelligence chip 11, the board 12 or the electronic device 13, the artificial intelligence chip 11, the board 12 or the electronic device 13 performs road surface detection in the target image, and outputs the road surface quality detection result through the display. As shown in Figure 1, in the road surface quality detection result, the road surface disease area is framed by a detection frame, and the category label of the road surface disease, or category text, etc. can be marked. For example, in Figure 1, the category label of the road surface disease is marked as D00, which is represented as a vertical crack.

[0049] It is understandable that the application scenario of the road quality detection method provided by the present application can also be to detect the road quality in real time. For example, the road quality collection vehicle can be equipped with a visual sensor, such as a high-definition camera, a laser radar, etc. The road video is collected by the visual sensor to obtain the target image in the road video, and the road quality detection device installed in the road quality collection vehicle or the road quality detection vehicle communicates with the cloud server to upload the road video or the target image to the cloud server, and the cloud server completes the detection of the road quality, and the image of the road detection result or the video after encoding the image of the road detection result can be stored in the cloud for the user to view.

[0050] It is understandable that the application scenario of the road surface quality detection method provided in this embodiment is not limited.

[0051] The following specific embodiments are used to describe in detail the technical solutions of the present invention and how the technical solutions of the present application solve the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.

[0052] Figure 3 This is a flow chart of a road surface quality detection method provided in an embodiment of the present application. The execution subject of the road surface quality detection method provided in this embodiment may be a road surface quality detection device. In practical applications, the road surface quality detection device may be implemented by a computer program, such as application software, etc., or by a medium storing relevant computer programs, such as a USB flash drive, a CD, a cloud, etc., or may also be implemented by a physical device and a virtual device integrated or installed with relevant computer programs, such as a chip, a board, etc.

[0053] In the embodiment of the present application, the road quality detection device may be located in an artificial intelligence chip, board or electronic device. The electronic device may be a digital computer in various forms, such as a cloud server, a cellular phone, a smart phone, a laptop computer, a desktop computer, a workbench, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers.

[0054] like Figure 3 As shown, the road surface quality detection method provided in this embodiment includes the following steps:

[0055] Step 101: Acquire a target image to be identified, where the target image includes a road surface area.

[0056] In this embodiment, a video including a road surface area may be stored in a database, and the video may be decoded to obtain a target image to be processed. Alternatively, in this embodiment, the target image may be directly stored in a database, and the target image to be identified may be obtained from the database.

[0057] Alternatively, in this embodiment, a video or image including the road surface area is collected by a visual sensor to obtain a target image to be identified.

[0058] It can be understood that, in the target image, all may be road areas, or may include not only road areas but also other objects, which is not limited in this embodiment.

[0059] Step 102, using the improved feature extraction network in the YOLOv5 model to perform feature extraction processing on the target image to obtain extracted feature data; the improved feature extraction network has a receptive field size suitable for road surface disease detection.

[0060] like Figure 4 As shown in the figure, the current YOLOv5 model includes an input image processing network, a feature extraction network, a feature enhancement network, and an output detection network.

[0061] Among them, the output image processing network includes a Mosaic data enhancement module and an adaptive image scaling module. Figure 4 The Mosaic Scale module is in the figure. The Mosaic data enhancement module can be used for the training process of the YOLOv5 model. Specifically, four images are randomly scaled, randomly cropped, randomly arranged, and spliced ​​after the color gamut is changed to obtain a sample image with richer content. The adaptive image scaling module is used in the inference process of road quality detection. The adaptive image scaling module is used to adaptively scale the image size to ensure consistency with the model input.

[0062] Among them, the feature extraction network also includes Focus modules and multiple CBL CSP1_Y modules. In the CBL CSP1_Y module. CBL is composed of Conv+Bn+Leaky_relu activation functions. In the CSP1_Y module, Y represents the number of residual components. The residual component can be expressed as Res unit. Among them, the main task of the Focus module is to perform slicing operations. The CBLCSP1_Y module and the CBL CSP1_Y module divide the feature map of the base layer into two parts, and then connect them through a cross-stage hierarchy to reduce the amount of calculation while ensuring accuracy. Among them Figure 4 P3, P4 and P5 in the figure represent the feature outputs corresponding to the 4th, 6th and 8th layers respectively.

[0063] The feature enhancement network includes: spatial pyramid pooling (referred to as SPP module), FPN, and path aggregation network (referred to as PAN), which further fuses and enhances the features extracted by the feature extraction network. Among them, FPN and PAN can be referred to as FPN+PAN.

[0064] Among them, the SPP module is used to perform multi-scale pooling on the input feature data and connect all the outputs to achieve the fusion of multi-scale feature information. FPN performs bottom-up feature fusion on the feature data output by the feature extraction network. PAN performs bottom-up feature enhancement on the features fused by FPN. After FPN+PAN processing, the features can be called: Feature Pyramid (English: Feature Pyramid).

[0065] In the output detection network, the feature map is first converted into a detection form. Then, each point of the feature map grid is calculated in combination with the preset prior box. When the confidence is greater than the preset confidence, the current prediction bbox information is retained. Finally, all the retained bboxes are combined and sent to the NMS model for re-screening to obtain the final prediction result. Figure 4 In the figure, we can see that in the final prediction results, there are vertical cracks D00, transverse cracks D10 and crocodile cracks D20.

[0066] Understandably, Figure 4 The YOLOv5 model in is an existing conventional model, and the algorithm implementation will not be described one by one.

[0067] In this embodiment, the feature extraction network in the YOLOv5 model is improved to obtain an improved feature extraction network. When improving the feature extraction network, by analyzing the size distribution of road surface diseases, it is known that each road surface disease has a different size distribution, but vertical cracks, horizontal cracks and crocodile patterns are all medium and large targets, and the feature extraction network must have enough receptive field to extract the characteristic data of the target. Therefore, when improving the feature extraction network, it is necessary to improve the receptive field size. The feature extraction network includes multiple feature output layers, each feature output layer corresponds to features of different sizes, and the corresponding receptive field size should also be adapted thereto. Therefore, a receptive field size suitable for road surface disease detection is set for each feature output layer.

[0068] In this embodiment, when determining the appropriate receptive field size for each feature output layer, it can be determined based on the target coverage rate, or determined by other methods, which is not limited in this embodiment.

[0069] Among them, the target coverage is the proportion of targets whose length and width of the output data are smaller than the corresponding length and width of the receptive field to the total targets.

[0070] Step 103, using the improved feature enhancement network in the YOLOv5 model to perform feature enhancement processing on the feature data to obtain enhanced features; the enhanced features are the features obtained after the feature data is initially fused and enhanced, and then the medium and large size features are fused and enhanced.

[0071] In this embodiment, the feature enhancement network in the YOLOv5 model is improved to obtain an improved feature enhancement network. When improving the feature enhancement network, the feature enhancement network in the YOLOv5 model is analyzed, and it is learned that the original feature enhancement network includes a network structure combining FPN and PAN, which is referred to as the FPN+PAN structure. In the FPN+PAN structure, the feature data output by the improved feature extraction network is subjected to bottom-up feature fusion and bottom-up feature enhancement. Among them, the feature fusion adopts the FPN network structure, and the feature enhancement adopts the PAN network structure. Therefore, the feature enhancement network uniformly fuses and enhances features of all sizes. The features after the feature fusion and enhancement processing of the feature data using the FPN+PAN structure can be determined as preliminary enhanced features.

[0072] However, pavement defects are more distributed in medium and large target areas, and it is necessary to further perform feature fusion and enhancement on medium and large size features. For example, a medium and large scale feature enhancement network structure can be used to further perform feature fusion and enhancement on medium and large size features.

[0073] Step 104: Detect road surface defects in the target image based on the enhanced features.

[0074] In this embodiment, the enhanced features are input into the output detection network in the YOLOv5 model, and the output detection network is used to detect pavement defects in the target image. If pavement defects exist in the target image, the pavement defects are framed with a detection frame, and the type label or type text of the pavement defect can be marked at the preset position of the frame.

[0075] The pavement quality detection method provided in the present embodiment obtains a target image to be identified, wherein the target image includes a pavement area; uses an improved feature extraction network in the YOLOv5 model to perform feature extraction processing on the target image to obtain extracted feature data; the improved feature extraction network has a receptive field size suitable for pavement disease detection; uses an improved feature enhancement network in the YOLOv5 model to perform feature enhancement processing on the feature data to obtain enhanced features; the enhanced features are features obtained by performing feature fusion enhancement processing on medium and large-sized features after preliminary fusion enhancement processing of the feature data; pavement diseases in the target image are detected based on the enhanced features, and since the improved feature extraction network has a receptive field size suitable for pavement disease detection, it has an optimal target coverage rate suitable for pavement diseases of different types and sizes, and can effectively extract pavement disease features. Since the improved feature enhancement network performs preliminary fusion and enhancement processing on the feature data, and then performs feature fusion and enhancement on the medium and large-sized features, it can effectively enhance the characteristics of medium and large-sized pavement diseases. By improving the feature extraction network and feature enhancement network in the YOLOv5 model, the improved YOLOv5 model can be more suitable for pavement quality detection scenarios, thereby effectively improving the accuracy of pavement quality detection.

[0076] Figure 5 This is a flow chart of a road quality detection method provided in another embodiment of the present application. This embodiment is an optional implementation method for determining an improved feature enhancement network solution in the YOLOv5 model based on the first embodiment. The road quality detection method provided in this embodiment also includes the following steps:

[0077] Step 201, setting the number parameters of CSP modules in multiple feature output layers.

[0078] In this embodiment, the YOLOv5 model has multiple feature output layers. Figure 4As shown, multiple feature output layers can be P3, P4 and P5. Among them, the features in P3 are used to detect small targets, the features in P4 are used to detect medium-sized targets, and the features in P5 are used to detect large targets. The receptive fields and target coverage rates of corresponding sizes in different feature output layers reflect the feature extraction network's ability to detect targets of that size. Therefore, it is necessary to improve the receptive field size and target coverage rate corresponding to each feature output layer to determine the most suitable receptive field size and the best target coverage rate. The receptive field size and target coverage rate are both related to the number parameters of the CSP modules in the feature output layer. Therefore, when determining the most suitable receptive field size and the best target coverage rate, it is necessary to first set the number parameters of the CSP modules in multiple feature output layers.

[0079] For example, if the feature output layers are P3, P4 and P5, the number parameters of the CSP modules may be set to N1, N2 and N3 respectively.

[0080] It should be noted that the number parameter of the CSP module also represents the number parameter of convolution added at the corresponding scale to adjust the receptive field.

[0081] Step 202, randomly select the number parameter corresponding to each feature output layer within a preset value range, and determine the corresponding candidate feature extraction network after each random selection.

[0082] In this embodiment, the preset value range may be {0, 1, 2, 3}, or may include a larger integer, which is not limited in this embodiment.

[0083] In this embodiment, the number parameter corresponding to each feature output layer can be randomly selected within a preset value range to form a parameter combination of the number of CSP modules corresponding to all feature output layers. Then, after multiple random selections, multiple parameter combinations are formed.

[0084] It can be understood that in order to fully cover all parameter combinations, the number parameters corresponding to all feature output layers can be traversed in combination within a preset value range, and each parameter combination can be traversed.

[0085] In this embodiment, the feature extraction network corresponding to each parameter combination is determined as a candidate feature extraction network, so as to determine an improved feature extraction network from multiple candidate feature extraction networks.

[0086] Step 203, determining the target coverage corresponding to each candidate feature extraction network.

[0087] As an optional implementation, step 203 includes the following steps:

[0088] Step 2031, calculating the receptive field size of each feature output layer of each candidate feature extraction network.

[0089] In this embodiment, when calculating the receptive field size of each feature output layer of each candidate feature extraction network, the receptive field calculation formula can be used for calculation. Specifically, the size of the convolution in the feature output layer and the size of the convolution kernel, as well as the jump size of the output feature map are obtained, and the size of the receptive field is calculated according to the size of the convolution in the feature output layer and the size of the convolution kernel, as well as the jump size of the output feature map.

[0090] Step 2032, calculating the coverage of different road surface diseases according to the receptive field size corresponding to each candidate feature extraction network.

[0091] In this embodiment, the coverage rate of different road surface diseases is calculated for each candidate feature extraction network.

[0092] Specifically, the receptive field size corresponding to each feature output layer in each candidate feature extraction network and the size of the input data corresponding to each feature output layer are obtained. The ratio of the size of the input data to the corresponding receptive field size can be calculated. If the ratio is less than or equal to 1, it means that the receptive field covers the target, otherwise it means that the receptive field does not cover the target, and then the coverage rate of each feature output layer for different pavement diseases is calculated. The average value of the coverage rate of different pavement diseases in each feature output layer can be calculated, or the parameter value that can comprehensively reflect the coverage rate of different pavement diseases in each feature output layer can be used to calculate the coverage rate of different pavement diseases.

[0093] Step 2033, determining the target coverage rate corresponding to each candidate feature extraction network according to the coverage rate of different road surface diseases corresponding to each candidate feature extraction network.

[0094] Specifically, the corresponding target coverage rate can be calculated for each candidate feature extraction network. Specifically, the average coverage rate of each candidate output layer for different pavement defects in each candidate feature extraction network or the parameter value that comprehensively reflects the coverage rate of different pavement defects in each feature output layer is obtained, and then the average coverage rate of each candidate output layer for different pavement defects can be averaged to determine the target coverage rate corresponding to the candidate feature extraction network.

[0095] In this embodiment, when determining the target coverage rate corresponding to each candidate feature extraction network, the receptive field size of each feature output layer of each candidate feature extraction network is calculated; the coverage rate of different road surface diseases is calculated based on the receptive field size corresponding to each candidate feature extraction network; and the target coverage rate corresponding to each candidate feature extraction network is determined based on the coverage rate of different road surface diseases corresponding to each candidate feature extraction network. Since the target coverage rate is the proportion of targets whose output data size is smaller than the receptive field size to the total targets, when calculating the target coverage rate corresponding to each candidate feature extraction network, the receptive field size of each feature output layer is first calculated, and the coverage rate of different road surface diseases and the target coverage rate corresponding to each candidate feature extraction network are calculated based on the receptive field size corresponding to each candidate feature extraction network, so that the target coverage rate corresponding to each candidate feature extraction network can be accurately calculated.

[0096] Step 204: determine a preset number of optimal target coverages from the target coverages corresponding to the candidate feature extraction networks.

[0097] As an optional implementation, in this embodiment, step 204 includes the following steps:

[0098] Step 2041, sort the target coverage rates corresponding to the candidate feature extraction networks in descending order.

[0099] Step 2042, determining a preset number of larger target coverages from the sorting results.

[0100] Specifically, in this embodiment, the target coverage rates corresponding to the candidate feature extraction networks can be sorted from large to small, and the first N target coverage rates sorted at the front can be determined as the optimal target coverage rates of the preset number.

[0101] Wherein, N is a preset number. The first one is the best target coverage rate. It comprehensively reflects the best coverage rate for targets of different sizes.

[0102] In this embodiment, when determining a preset number of optimal target coverage rates from the target coverage rates corresponding to each candidate feature extraction network, the target coverage rates corresponding to each candidate feature extraction network are sorted in order from large to small, and the preset number of larger target coverage rates are determined from the sorting results, so that the preset number of larger target coverage rates can be determined quickly and accurately.

[0103] Step 205 , selecting an improved feature extraction network from a preset number of candidate feature extraction networks corresponding to better target coverage.

[0104] As an optional implementation, in this embodiment, step 205 includes the following steps:

[0105] Step 2051, determining the parameter amount and computation amount of the candidate feature extraction network corresponding to a preset number of optimal target coverage rates.

[0106] In this embodiment, the operation process of the N candidate feature extraction networks is monitored to determine the parameter quantities and calculation quantities corresponding to the N candidate extraction networks.

[0107] Step 2052, selecting an improved feature extraction network from a preset number of candidate feature extraction networks according to a preset number of preferred target coverages, corresponding parameter quantities, and calculation quantities.

[0108] In this embodiment, N candidate feature extraction networks are compared from the dimensions of target coverage, parameter quantity and calculation amount. If there is a candidate feature extraction network with the largest target coverage, the smallest parameter quantity and the smallest calculation amount, the candidate feature extraction network with the largest target coverage, the smallest parameter quantity and the smallest calculation amount is determined as the improved feature extraction network. If there is no candidate feature extraction network with the largest target coverage, the smallest parameter quantity and the smallest calculation amount, the target coverage, parameter quantity and calculation amount can be normalized, and the corresponding weights can be set, and the weighted sum result can be calculated by weighted summing the target coverage, parameter quantity and calculation amount. The candidate feature extraction network with the largest weighted sum result is determined as the improved feature extraction network.

[0109] In this embodiment, when selecting an improved feature extraction network from the candidate feature extraction networks corresponding to the preset number of better target coverage rates, the parameter amount and calculation amount of the candidate feature extraction networks corresponding to the preset number of better target coverage rates are determined; the improved feature extraction network is selected from the preset number of candidate feature extraction networks according to the preset number of better target coverage rates, the corresponding parameter amount and calculation amount, and when determining the improved feature extraction network, the three factors of target coverage rate, parameter amount and calculation amount can be comprehensively considered, so that the determined improved feature extraction network not only has a better target coverage rate, but also has a smaller parameter amount and calculation amount. In this way, the accuracy of road quality detection is guaranteed while improving the efficiency of road quality detection.

[0110] In this embodiment, a candidate extraction network corresponding to the best target coverage is obtained. The candidate extraction network corresponding to the best target coverage has a receptive field distribution that is most suitable for road disease detection, so the candidate extraction network corresponding to the best target coverage is determined as an improved feature extraction network.

[0111] The pavement quality detection method provided in the present embodiment sets the number parameters of CSP modules in multiple feature output layers when determining the improved feature enhancement network in the YOLOv5 model; randomly takes values ​​of the number parameters corresponding to each feature output layer within a preset value range, and determines the corresponding candidate feature extraction network after each random value taking; determines the target coverage rate corresponding to each candidate feature extraction network; determines a preset number of better target coverage rates from the target coverage rates corresponding to each candidate feature extraction network; selects an improved feature extraction network from the candidate feature extraction networks corresponding to the preset number of better target coverage rates, determines multiple candidate feature extraction networks by setting the number parameters of CSP modules in the feature output layer related to both the receptive field size and the target coverage rate, and selects the candidate feature extraction network corresponding to the better target coverage rate as the improved feature extraction network by calculating the target coverage rate corresponding to each candidate feature extraction network, so that the improved feature extraction network can have the most suitable receptive field distribution for pavement disease detection, and can fully extract the characteristics of different pavement diseases, thereby effectively improving the ability to extract pavement disease characteristics.

[0112] Figure 6 This is a flow chart of a road quality detection method provided by another embodiment of the present application. This embodiment is an optional implementation method of randomly selecting the number parameters corresponding to each feature output layer within a preset value range based on the above-mentioned embodiment. The road quality detection method provided by this embodiment also includes the following steps:

[0113] Step 301, split the number parameters of the CSP modules in each feature output layer into length dimension and width dimension to obtain the length number parameters and width number parameters corresponding to each feature output layer.

[0114] Step 302: for each feature output layer, the length number parameter and the width number parameter are combined to form a number parameter combination corresponding to each feature output layer.

[0115] In this embodiment, the receptive field can be a matrix data, and the input data can also be a matrix data, so when considering the influence of the number parameter of the CSP modules in each feature output layer on the receptive field size, it is necessary to consider not only the influence on the overall scale of the receptive field, but also the influence on the receptive field size under different dimensions of the same scale. Therefore, the number parameter of the CSP modules in each feature output layer is split into the length dimension and the width dimension.

[0116] Among them, if the number parameter of the CSP module in each feature output layer is represented as Ni, then the number parameter of the CSP module in each feature output layer is split into the length dimension and the width dimension, and the length number parameter and the width number parameter are combined to form the number parameter combination corresponding to each feature output layer, which can be expressed as <ni w ,ni h >.

[0117] For example, when the multiple feature output layers are P3, P4 and P5, the number parameters of the CSP modules are set to N1, N2 and N3. Then, after the number parameters of the CSP modules in each feature output layer are split into the length dimension (H dimension) and the width dimension (W dimension), the length number parameters and width number parameters corresponding to the formed P3, P4 and P5 feature output layers are respectively <n1 w ,n1 h >, <n2 w ,n2 h >, <n3 w ,n3 h >.

[0118] Step 303: randomly select values ​​of the length number parameter and the width number parameter in the number parameter combination corresponding to each feature output layer within a preset value range.

[0119] In this embodiment, the preset value range may be {0, 1, 2, 3}, or may include a larger integer, which is not limited in this embodiment.

[0120] In this embodiment, the length number parameter and the width number parameter in the number parameter combination corresponding to each feature output layer are randomly selected within the preset value range. In order to fully cover all parameter combinations, the length number parameter and the width number parameter corresponding to all feature output layers can be combined and traversed within the preset value range, and each parameter combination is traversed.

[0121] The road quality detection method provided by the present embodiment, when randomly taking values ​​of the number parameters corresponding to each feature output layer within a preset value range, splits the number parameters of the CSP modules in each feature output layer into length dimension and width dimension to obtain the length number parameters and width number parameters corresponding to each feature output layer; for each feature output layer, combines the length number parameters and the width number parameters to form a number parameter combination corresponding to each feature output layer; randomly takes values ​​of the length number parameters and the width number parameters in the number parameter combination corresponding to each feature output layer within a preset value range. Since the number parameters of the CSP modules in each feature output layer are split into length dimension and width dimension, and then the length number parameters and the width number parameters in the number parameter combination corresponding to each feature output layer are randomly taken within a preset value range, the number parameters of the CSP modules in each feature output layer can not only reflect the influence on the receptive fields of different scales but also reflect the influence on the receptive fields of different dimensions at the same scale.

[0122] Figure 7 A flow chart of a road quality detection method provided in another embodiment of the present application is as follows: Figure 7 As shown, the road quality detection method provided in this embodiment is based on the road quality detection method provided in any of the above embodiments, and uses the improved feature extraction network in the YOLOv5 model to perform feature extraction processing on the target image to obtain further refinement of the extracted feature data.

[0123] Specifically, Figure 8 As shown, in the improved YOLOv5 model provided in this embodiment, the improved feature extraction network includes multiple feature extraction CSP modules, and a channel attention mechanism ECAM module is connected after each CSP module. Figure 8 The ECA-CSP in the embodiment of the present application is a CSP combined with an ECAM module.

[0124] In the previous embodiment, the improved feature extraction network has better target coverage. Correspondingly, the number of CSP models is increased for each feature output layer to adjust the receptive field size of the corresponding scale, so that the target feature data can be extracted more accurately. However, since the intermediate feature map extracted from the feature extraction network often has multiple channels, and the feature maps of different channels express different feature information, it is necessary to screen the important feature information and remove invalid feature information. Therefore, based on the CSP model itself, the channel attention mechanism is used to improve it, that is, a channel attention mechanism ECAM module is connected after each CSP module. Through the CSP module and the connected ECAM module, important features are identified while extracting features, so that important features are obtained and invalid features are weakened.

[0125] In the road quality detection method provided in this embodiment, the improved feature extraction network in the YOLOv5 model is used to perform feature extraction processing on the target image to obtain the extracted feature data, which includes the following steps:

[0126] Step 401: Use the CSP module to extract the features of the target image to obtain initial feature data.

[0127] The initial feature data may be represented in the form of a feature map.

[0128] As an optional implementation, step 401 includes the following steps:

[0129] Step 401a, using the main branch of the CSP module to perform multiple convolution operations, and using the secondary branch of the CSP module to perform residual connection.

[0130] Step 401b, using the concatenation operator Concat to concatenate the main branch processing result with the secondary branch processing result to obtain initial feature data.

[0131] Specifically, Fig.10 As shown in FIG. 1 , in the CSP module, there are multiple CBL submodules on the main branch. The CBL submodule has multiple convolution operations, which can expand the receptive field and extract feature information. The CBL submodule on the secondary branch is used for residual connection. The processing results of the main branch and the secondary branch are concatenated by the concatenation operator Concat( Fig.10 The spliced ​​features are processed by a CBL submodule to obtain the initial feature data.

[0132] Among them, Fig.10 In it, x=1 and y is the number of residual components.

[0133] Step 402: Use the ECAM module to screen important features of the initial feature data to obtain extracted feature data.

[0134] As an optional implementation, step 402 includes the following steps:

[0135] Step 402a, using the ECAM module to calculate the weight vector of the initial feature data channel dimension.

[0136] Step 402b, filtering important feature data in the initial feature data according to the weight vector of the channel dimension.

[0137] Step 402c: determining the important feature data as the extracted feature data.

[0138] Specifically, in this embodiment, the structure of the channel attention mechanism ECAM module is as follows: Fig. 9As shown in the figure, the ECAM module first uses global average pooling (Global AvgPooling, abbreviated as GAP) to pool the initial feature data to obtain the feature mean of each initial feature data, and the result is represented as a one-dimensional vector V. Then, one-dimensional convolution (Conv1d for short) is used to perform vector feature fusion processing on V, where the convolution kernel size K can be adjusted according to the number of channels by nonlinear mapping, and the result is represented as a one-dimensional vector V'. Then, the Sigmoid activation function is used to activate V' to obtain the weight vector V". Finally, V" is used to weight the initial feature data from the channel dimension (C dimension for short) to complete the channel attention operation, so as to realize the important feature screening of the initial feature data, so as to obtain the extracted feature data δ'.

[0139] The initial feature data is represented as δ, and its dimension is H*W*C. The convolution kernel size K can be 5.

[0140] It should be noted that, in the ECAM module of this embodiment, Conv1d is used to replace the traditional MLP structure for vector feature fusion. Since Conv1d has a smaller amount of calculation and parameters, and replaces the full connection with a local connection, it not only reduces the complexity of the ECAM module, but also avoids degradation, which is of great significance for maintaining the integrity of channel information.

[0141] The pavement quality detection method provided in this embodiment uses a CSP module to extract features of the target image to obtain initial feature data when using the improved feature extraction network in the YOLOv5 model to perform feature extraction processing on the target image to obtain extracted feature data; and uses an ECAM module to screen important features of the initial feature data to obtain extracted feature data. Since the ECAM module has an efficient channel attention mechanism, the ECAM module can effectively screen important features of the initial feature data and weaken invalid features, so that the improved feature extraction network can more effectively extract important features of pavement diseases.

[0142] Fig.16 A flow chart of a road quality detection method provided in another embodiment of the present application is as follows: Fig.16 As shown, the road surface quality detection method provided in this embodiment is based on the road surface quality detection method provided in any of the above embodiments, and further refines the feature data by using the improved feature enhancement network in the YOLOv5 model to obtain the enhanced feature scheme, and the road surface quality detection method provided in this embodiment also includes the following steps:

[0143] Step 501, using a network structure combining FPN and PAN to perform preliminary feature fusion and enhancement processing on feature data to obtain preliminary enhanced features.

[0144] like Figure 8 and Fig.12 As shown, in this embodiment, the network structure combining FPN and PAN is referred to as FPN+PAN structure.

[0145] Specifically, the FPN+PAN structure is used to perform bottom-up feature fusion and bottom-up feature enhancement on the feature data output by the improved feature extraction network to obtain preliminary enhanced features, wherein the preliminary enhanced features can be feature pyramids.

[0146] For example, in the YOLOv5S model, the scales of the three-layer feature pyramid are: P3<128,96,96>, P4<256,48,48> and P5<512,24,24>.

[0147] Step 502, using a medium- and large-scale feature enhancement network structure to perform feature fusion enhancement processing on the medium- and large-scale features based on the preliminary enhanced features, so as to obtain medium- and large-scale enhanced features.

[0148] In this embodiment, since road surface disease targets are more distributed in medium and large target areas, and among multiple feature output layers, P3 is mainly used to detect small targets, and P4 and P5 are used to detect medium and large targets, a medium and large scale feature enhancement network structure is used to perform feature fusion enhancement processing on medium and large scale features based on preliminary enhanced features to obtain medium and large scale enhanced features.

[0149] Optionally, the medium- and large-scale feature enhancement network structure includes: a network structure combining multiple FPNs and PANs. Fig.12 The medium and large scale feature enhancement network structure can be expressed as N Sin-FPN Blocks.

[0150] As an optional implementation, step 502 includes the following steps:

[0151] Step 502a, inputting the medium-scale feature data and the large-scale feature data into the large-scale feature enhancement network structure.

[0152] Exemplarily, the data output by the FPN+PAN structure corresponding to the feature output layer of P4 is used as the medium-scale feature data, and the data output by the FPN+PAN structure corresponding to the feature output layer of P5 is used as the large-scale feature data. The medium-scale feature data and the large-scale feature data are first input into the first FPN and PAN combined network structure in the medium- and large-scale feature enhanced network structure.

[0153] Step 502b, using the first FPN and PAN combined network structure to perform feature fusion and enhancement processing on the medium-scale feature data and the large-scale feature data to obtain fusion and enhanced features.

[0154] Specifically, in this embodiment, the FPN in the first network structure combining FPN and PAN is used to perform feature fusion processing on the medium-scale feature data and the large-scale feature data, and then the feature fusion processing results and the large-scale feature data are subjected to feature enhancement processing using the PAN in the first network structure combining FPN and PAN to obtain the fusion and enhanced features of the first network structure combining FPN and PAN.

[0155] Step 502c, using at least one FPN and PAN combined network structure in the middle position to perform feature fusion and enhancement processing on the previous fusion and enhancement features in turn, and the last FPN and PAN combined network structure outputs medium and large scale enhanced features.

[0156] Specifically, the fusion enhanced features output by the first FPN and PAN combined network structure and the feature fusion processing results output by the FPN are input into the second FPN and PAN combined network structure in the medium and large scale feature enhanced network structure. First, the FPN in the second FPN and PAN combined network structure is used for feature fusion processing again, and then the feature fusion processing results and the fusion enhanced features output by the first FPN and PAN combined network structure are used for feature enhancement processing again to obtain the fusion enhanced features of the second FPN and PAN combined network structure. If there are multiple FPN and PAN combined network structures in the middle position, the third FPN and PAN combined network structure is used to perform feature fusion enhancement processing on the previous fusion enhanced features again, and so on, until the medium and large scale enhanced features are output by the last FPN and PAN combined network structure.

[0157] Among them, the medium and large scale enhanced features output by the last network structure combining FPN and PAN include: the feature fusion processing results and feature enhancement processing results corresponding to the last network structure combining FPN and PAN.

[0158] It should be noted that in Figure 8 In the feature output layers P4 and P5, the features are further fused and strengthened. At this time, there are only two layers, P4 and P5, so it is 2*N.

[0159] In this embodiment, when a medium- and large-scale feature enhancement network structure is used to perform feature fusion enhancement processing on medium- and large-scale features based on preliminary enhancement features to obtain medium- and large-scale enhanced features, the medium-scale feature data and the large-scale feature data are input into the medium- and large-scale feature enhancement network structure; the first FPN and PAN combined network structure is used to perform feature fusion enhancement processing on the medium- and large-scale feature data to obtain fusion enhanced features; at least one FPN and PAN combined network structure in the middle position is used to perform feature fusion enhancement processing on the previous fusion enhanced feature in turn, and the last FPN and PAN combined network structure outputs the medium- and large-scale enhanced features; multiple FPN and PAN combined network structures are used to perform feature fusion enhancement processing on medium- and large-scale features, which can effectively improve the feature enhancement effect of medium- and large-scale pavement disease features.

[0160] Step 503, determining enhanced features based on the preliminary enhanced features and the medium and large scale enhanced features.

[0161] In this embodiment, the fusion result of the small-scale features in the preliminary enhanced features and the large-scale enhanced features in the last network structure output combining FPN and PAN are determined as the enhanced features.

[0162] The pavement quality detection method provided in this embodiment uses the improved feature enhancement network in the YOLOv5 model to perform feature enhancement processing on feature data to obtain enhanced features, and then uses a network structure combining FPN and PAN to perform preliminary feature fusion enhancement processing on feature data to obtain preliminary enhanced features; uses a medium-to-large scale feature enhancement network structure to perform feature fusion enhancement processing on medium-to-large scale features based on preliminary enhanced features to obtain medium-to-large scale enhanced features; and determines enhanced features based on preliminary enhanced features and medium-to-large scale enhanced features. Since pavement disease targets are more distributed in medium-to-large target areas, after using a network structure combining FPN and PAN to perform feature fusion enhancement processing on pavement disease features of different sizes, and then uses a medium-to-large scale feature enhancement network structure to perform feature fusion enhancement processing on medium-to-large scale features based on preliminary enhanced features, it is possible to more fully enhance the features of medium-to-large scale pavement disease features, making the enhanced features more suitable for pavement disease targets.

[0163] It should be noted that, in the road quality detection methods provided in the above embodiments, the training process of the improved YOLOv5 model is similar to the training process of the existing YOLOv5 model, and will not be described again one by one.

[0164] Fig.13 A schematic diagram of a road surface quality detection device provided in one embodiment of the present application is shown in FIG. Fig.13As shown, the road quality detection device 60 provided in this embodiment includes: an acquisition module 61, a feature extraction module 62, a feature enhancement module 63 and a detection module 64.

[0165] Among them, the acquisition module 61 is used to acquire the target image to be identified, and the target image includes the road surface area. The feature extraction module 62 is used to use the improved feature extraction network in the YOLOv5 model to perform feature extraction processing on the target image to obtain extracted feature data; the improved feature extraction network has a receptive field size suitable for road surface disease detection. The feature enhancement module 63 is used to use the improved feature enhancement network in the YOLOv5 model to perform feature enhancement processing on the feature data to obtain enhanced features; the enhanced features are the features obtained after the feature data is initially fused and enhanced, and then the medium and large size features are fused and enhanced. The detection module 64 is used to detect road surface diseases in the target image based on the enhanced features.

[0166] Optionally, the improved feature extraction network includes a plurality of feature output layers;

[0167] Correspondingly, the road surface quality detection device also includes: a first network determination module.

[0168] Among them, the first network determination module is used to set the number parameters of CSP modules in multiple feature output layers; randomly select the number parameters corresponding to each feature output layer within a preset value range, and determine the corresponding candidate feature extraction network after each random selection; determine the target coverage rate corresponding to each candidate feature extraction network; determine a preset number of better target coverage rates from the target coverage rates corresponding to each candidate feature extraction network; and screen out an improved feature extraction network from the candidate feature extraction networks corresponding to the preset number of better target coverage rates.

[0169] Optionally, when the first network determination module randomly selects a value for the number parameter corresponding to each feature output layer within a preset value range, it is specifically used to:

[0170] The number parameters of the CSP modules in each feature output layer are split into length dimension and width dimension to obtain the length number parameters and width number parameters corresponding to each feature output layer; for each feature output layer, the length number parameters and the width number parameters are combined to form the number parameter combination corresponding to each feature output layer; the length number parameters and the width number parameters in the number parameter combination corresponding to each feature output layer are randomly selected within a preset value range.

[0171] Optionally, when determining the target coverage corresponding to each candidate feature extraction network, the first network determination module is specifically used to:

[0172] Calculate the receptive field size of each feature output layer of each candidate feature extraction network; calculate the coverage rate of different pavement defects according to the receptive field size corresponding to each candidate feature extraction network; determine the target coverage rate corresponding to each candidate feature extraction network according to the coverage rate of different pavement defects corresponding to each candidate feature extraction network.

[0173] Optionally, when the first network determination module selects the improved feature extraction network from the candidate feature extraction networks corresponding to a preset number of better target coverage rates, it is specifically used to:

[0174] Determine the parameter amount and calculation amount of the candidate feature extraction network corresponding to the preset number of better target coverage rates; and select an improved feature extraction network from the preset number of candidate feature extraction networks based on the preset number of better target coverage rates, the corresponding parameter amount and the calculation amount.

[0175] Optionally, the improved feature extraction network includes: a plurality of feature extraction CSP modules, and a channel attention mechanism ECAM module is connected after each CSP module;

[0176] Accordingly, the feature extraction module 62 is specifically used for:

[0177] The CSP module is used to extract the features of the target image to obtain initial feature data; the ECAM module is used to screen important features of the initial feature data to obtain extracted feature data.

[0178] Optionally, the feature extraction module 62, when using the CSP module to extract the features of the target image to obtain initial feature data, is specifically used to:

[0179] The main branch of the CSP module is used to perform multiple convolution operations, and the secondary branch of the CSP module is used for residual connection. The concatenation operator Concat is used to concatenate the main branch processing result with the secondary branch processing result to obtain the initial feature data.

[0180] Optionally, the feature extraction module 62, when using the ECAM module to screen the initial feature data for important features to obtain the extracted feature data, is specifically used to:

[0181] The ECAM module is used to calculate the weight vector of the channel dimension of the initial feature data; the important feature data in the initial feature data is screened according to the weight vector of the channel dimension;

[0182] The important feature data are determined as the extracted feature data.

[0183] Optionally, the improved feature enhancement network includes a network structure combining FPN and PAN, and a medium- and large-scale feature enhancement network structure.

[0184] Accordingly, the feature enhancement module 63 is specifically used for:

[0185] A network structure combining FPN and PAN is used to perform preliminary feature fusion and enhancement processing on feature data to obtain preliminary enhanced features; a medium and large scale feature enhancement network structure is used to perform feature fusion and enhancement processing on medium and large scale features based on preliminary enhanced features to obtain medium and large scale enhanced features; the enhanced features are determined based on the preliminary enhanced features and medium and large scale enhanced features.

[0186] Optionally, the medium- and large-scale feature enhancement network structure includes: a network structure combining multiple FPNs and PANs;

[0187] The feature enhancement module 63, when using the medium-to-large scale feature enhancement network structure to perform feature fusion enhancement processing on the medium-to-large scale features based on the preliminary enhanced features to obtain the medium-to-large scale enhanced features, is specifically used to:

[0188] The medium-scale feature data and the large-scale feature data are input into the medium- and large-scale feature enhancement network structure; the first FPN and PAN combined network structure is used to perform feature fusion and enhancement processing on the medium-scale feature data and the large-scale feature data to obtain fusion and enhancement features; at least one FPN and PAN combined network structure in the middle position is used to perform feature fusion and enhancement processing on the previous fusion and enhancement features in turn, and the last FPN and PAN combined network structure outputs the medium- and large-scale enhanced features.

[0189] The road surface quality detection device provided in this embodiment can execute the road surface quality detection method provided in any of the above embodiments, and the specific principles and beneficial effects will not be repeated here.

[0190] Fig.14 A schematic diagram of a road surface quality detection device provided in another embodiment of the present application is shown in FIG. Fig.14 As shown, the present application embodiment provides a road quality detection device 70, including: a processor 72 and a memory 71; wherein the memory 71 is used to store program codes; the processor 72 is used to call the program codes stored in the memory to execute the road quality detection method provided by any of the above embodiments. The relevant description can be understood by referring to the relevant descriptions and effects corresponding to the steps in the accompanying drawings, and no further elaboration is made here.

[0191] The program may include program code, and the program code includes computer-executable instructions. The memory 801 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory.

[0192] In this embodiment, the memory 71 and the processor 72 are connected via a bus. The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Fig.14 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0193] An embodiment of the present application also provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a computer, the computer executes the road quality detection method provided by any one of the above embodiments.

[0194] An embodiment of the present application also provides a computer program product, including a computer program, characterized in that when the computer program is executed by a processor, the road surface quality detection method provided by any one of the above embodiments is implemented.

[0195] The embodiments of the present application also provide an artificial intelligence chip and board. Fig.15 is a structural diagram showing a board according to an embodiment of the present application. Fig.15 As shown, the board 80 includes a chip 801, which is a system-on-chip (SoC), or system on chip, integrated with one or more combined processing devices, which are artificial intelligence computing units that support various deep learning and machine learning algorithms to meet the intelligent processing needs in complex scenarios in the fields of computer vision, speech, natural language processing, data mining, etc. In particular, deep learning technology is widely used in the field of cloud intelligence. A notable feature of cloud intelligence applications is the large amount of input data, which places high demands on the storage capacity and computing power of the platform. The board 80 of this embodiment is suitable for cloud intelligence applications and has huge off-chip storage, on-chip storage, and powerful computing power.

[0196] The chip 801 is connected to an external device 803 via an external interface device 802. The external device 803 is, for example, a server, a computer, a camera, a display, a mouse, a keyboard, a network card, or a wifi interface. The data to be processed can be transmitted from the external device 803 to the chip 801 via the external interface device 802. The calculation result of the chip 801 can be transmitted back to the external device 803 via the external interface device 802. According to different application scenarios, the external interface device 802 can have different interface forms, such as a PCIe interface.

[0197] The board 80 also includes a storage device 804 for storing data, which includes one or more storage units 805. The storage device 804 is connected to the control device 806 and the chip 801 through a bus and transmits data. The control device 806 in the board 80 is configured to regulate the state of the chip 801. To this end, in an application scenario, the control device 806 may include a single-chip microcomputer (Micro Controller Unit, MCU).

[0198] In a possible implementation, a combined processing device is also provided. Fig.16 8 is a block diagram showing the combined processing device in the chip 801 of this embodiment. Fig.16 As shown in , the combined processing device 90 includes a computing device 901 , an interface device 902 , a processing device 903 and a storage device 904 .

[0199] The computing device 901 is configured to perform user-specified operations, and is mainly implemented as a single-core intelligent processor or a multi-core intelligent processor to perform deep learning or machine learning calculations. It can interact with the processing device 903 through the interface device 902 to jointly complete the user-specified operations.

[0200] The interface device 902 is used to transmit data and control instructions between the computing device 901 and the processing device 903. For example, the computing device 901 can obtain input data from the processing device 903 via the interface device 902 and write it into the storage device on the computing device 901 chip. Further, the computing device 901 can obtain control instructions from the processing device 903 via the interface device 902 and write them into the control cache on the computing device 901 chip. Alternatively or optionally, the interface device 902 can also read data in the storage device of the computing device 901 and transmit it to the processing device 903.

[0201] The processing device 903, as a general processing device, performs basic controls including but not limited to data handling, starting and / or stopping the computing device 901. Depending on the implementation, the processing device 903 can be a central processing unit (CPU), a graphics processing unit (GPU) or one or more types of processors in other general and / or special processors, these processors include but are not limited to digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., and their number can be determined according to actual needs. As before, only with respect to the computing device 901 of the present application, it can be regarded as having a single-core structure or a homogeneous multi-core structure. However, when the computing device 901 and the processing device 903 are integrated and considered together, the two are regarded as forming a heterogeneous multi-core structure.

[0202] The storage device 904 is used to store data to be processed, which may be DRAM 904 , which is a DDR memory, and is usually 16G or larger in size, and is used to store data of the computing device 901 and / or the processing device 903 .

[0203] Fig.17 , which is a schematic diagram of the internal structure of a single-core computing device. The single-core computing device 1000 is used to process input data such as computer vision, speech, natural language, and data mining. The single-core computing device 1000 includes three modules: a control module 1001, a computing module 1002, and a storage module 1003.

[0204] The control module 1001 is used to coordinate and control the operation of the computing module 1002 and the storage module 1003 to complete the deep learning task, and includes an instruction fetch unit (IFU) 10011 and an instruction decode unit (IDU) 10012. The instruction fetch unit 10011 is used to obtain instructions from the processing device, and the instruction decode unit 10012 decodes the obtained instructions and sends the decoding results to the computing module 1002 and the storage module 1003 as control information.

[0205] The operation module 1002 includes a vector operation unit 10021 and a matrix operation unit 10022. The vector operation unit 10021 is used to perform vector operations and can support complex operations such as vector multiplication, addition, and nonlinear transformation; the matrix operation unit 10022 is responsible for the core calculation of the deep learning algorithm, namely matrix multiplication and convolution.

[0206] The storage module 1003 is used to store or transfer relevant data, including a neuron RAM (NRAM) 10031, a parameter RAM (WRAM) 10032, and a direct memory access module (DMA) 10033. NRAM 10031 is used to store input neurons, output neurons, and intermediate results after calculation; WRAM 10032 is used to store the convolution kernel of the deep learning network, that is, the weight; DMA 10033 is connected to DRAM 904 through bus 1004, and is responsible for data transfer between the single-core computing device 1000 and DRAM 904.

[0207] Fig.18 , a schematic diagram of the internal structure of a computing device with multiple cores is shown. The multi-core computing device 1100 adopts a hierarchical structure design. The multi-core computing device 1100 is a system on chip, which includes at least one cluster, and each cluster includes multiple processor cores. In other words, the multi-core computing device 1100 is composed of a hierarchy of system on chip-cluster-processor core.

[0208] At the system-on-chip level, Fig.18 As shown, the multi-core computing device 1100 includes an external storage controller 1101 , a peripheral communication module 1102 , an on-chip interconnect module 1103 , a synchronization module 1104 and multiple clusters 1105 .

[0209] There may be multiple external storage controllers 1101, and two are shown in the figure as an example, which are used to respond to access requests issued by the processor core and access external storage devices, such as Fig.16DRAM 904 in the chip, thereby reading data from outside the chip or writing data. The peripheral communication module 1102 is used to receive control signals from the processing device through the interface device 902 to start the computing device 900 to perform tasks. The on-chip interconnect module 1103 connects the external storage controller 1101, the peripheral communication module 1102 and multiple clusters 1105 to transmit data and control signals between each module. The synchronization module 1104 is a global synchronization barrier controller (GBC) used to coordinate the work progress of each cluster and ensure information synchronization. Multiple clusters 1105 are the computing cores of the multi-core computing device 1100. Four are shown in the figure as an example. The four clusters 1105 form four quadrants as shown in Figure 1. With the development of hardware, the multi-core computing device 1101 of the present application can also include 8, 16, 64, or even more clusters 1105. Clusters 1105 are used to efficiently execute deep learning algorithms.

[0210] At the cluster level, Fig.18 As shown, each cluster 1105 includes multiple processor cores (IPU cores) 1106 and a memory core (MEM core) 1107. Exemplarily, each cluster 1105 includes 4 processor cores and 1 memory, and the memory may be DRAM 904.

[0211] The figure shows four processor cores 1106 as an example, but the present application does not limit the number of processor cores 1106. Fig.16 Each processor core 1106 is similar to Fig.19The single-core computing device 1200 also includes three modules: a control module 1201, a computing module 1202 and a storage module 1203. The functions and structures of the control module 1201, the computing module 1202 and the storage module 1203 are roughly the same as those of the control module 1001, the computing module 1002 and the storage module 1003. The control module 1201 includes an instruction fetch unit 12011 and an instruction decoding unit 12012. The computing module 1202 includes a vector computing unit 12021 and a matrix computing unit 12022. No further details are given. It should be noted that the storage module 1203 includes an input / output direct memory access module (IODMA) 12033 and a move direct memory access module (MVDMA) 12034. IODMA 12033 controls memory access between NRAM 12031 / WRAM 12032 and DRAM 904 via broadcast bus 1109 ; MVDMA 12034 is used to control memory access between NRAM 12031 / WRAM 12032 and storage unit (SRAM) 1108 .

[0212] Back to Fig.18 The storage core 1107 is mainly used for storage and communication, that is, to store shared data or intermediate results between the processor cores 1106, and to perform communication between the cluster 1105 and the DRAM 904, communication between the clusters 1105, and communication between the processor cores 1106. In other embodiments, the storage core 1107 has the ability of scalar operations and is used to perform scalar operations.

[0213] The storage core 1107 includes an SRAM 1108, a broadcast bus 1109, a cluster direct memory access module (cluster direct memory access, CDMA) 1110 and a global direct memory access module (global direct memory access, GDMA) 1111. The SRAM 1108 plays the role of a high-performance data transfer station. The data reused between different processor cores 1106 in the same cluster 1105 does not need to be obtained from the DRAM 904 by each processor core 1106, but is transferred between the processor cores 1106 via the SRAM 1108. The storage core 1107 only needs to quickly distribute the reused data from the SRAM 1108 to multiple processor cores 1106, so as to improve the efficiency of inter-core communication and greatly reduce on-chip and off-chip input / output access.

[0214] The broadcast bus 1109, CDMA 910 and GDMA 1111 are respectively used to perform communication between processor cores 1106, communication between clusters 1105 and data transmission between clusters 1105 and DRAM 904. They will be described below.

[0215] The broadcast bus 1109 is used to complete high-speed communication between the processor cores 1106 in the cluster 1105. The broadcast bus 1109 of this embodiment supports inter-core communication modes including unicast, multicast and broadcast. Unicast refers to point-to-point (e.g., single processor core to single processor core) data transmission, multicast is a communication mode of transmitting a copy of data from SRAM 1108 to specific processor cores 1106, and broadcast is a communication mode of transmitting a copy of data from SRAM 1108 to all processor cores 1106, which is a special case of multicast.

[0216] CDMA 1110 is used to control memory access to SRAM 1108 between different clusters 1105 within the same computing device.

[0217] GDMA 1111 cooperates with external memory controller 1101 to control the access of SRAM 1108 of cluster 1105 to DRAM 904, or reads data from DRAM 904 to SRAM 1108. As can be seen from the above, the communication between DRAM 904 and NRAM 10031 or WRAM 10032 can be realized through two channels. The first channel is to directly connect DRAM 904 and NRAM 10031 or WRAM 10032 through IODAM 10033; the second channel is to first transmit data between DRAM 904 and SRAM 1108 through GDMA 1111, and then transmit data between SRAM 1108 and NRAM 10031 or WRAM 10032 through MVDMA 10034. Although it seems that the second channel requires more components to participate and the data flow is longer, in fact, in some embodiments, the bandwidth of the second channel is much greater than the first channel, so the communication between DRAM 904 and NRAM 10031 or WRAM 10032 may be more efficient through the second channel. The embodiments of the present application can select the data transmission channel according to their own hardware conditions.

[0218] In other embodiments, the functions of GDMA 1111 and IODMA 10033 can be integrated into the same component. For the convenience of description, this application regards GDMA 1111 and IODMA 10033 as different components. For those skilled in the art, as long as the functions implemented and the technical effects achieved are similar to those of this application, they belong to the protection scope of this application. Furthermore, the functions of GDMA 1111, IODMA 10033, CDMA 1110, and MVDMA 10034 can also be implemented by the same component.

[0219] An embodiment of the present application also provides an electronic device, which includes an artificial intelligence chip or board.

[0220] The foregoing may be better understood in view of the following clauses:

[0221] Clause 1. A method for detecting road surface quality, comprising:

[0222] Acquire a target image to be identified, wherein the target image includes a road surface area;

[0223] Using an improved feature extraction network in a YOLOv5 model to perform feature extraction processing on the target image to obtain extracted feature data; the improved feature extraction network has a receptive field size suitable for pavement disease detection;

[0224] The feature data is subjected to feature enhancement processing by using the improved feature enhancement network in the YOLOv5 model to obtain enhanced features; the enhanced features are the features obtained by performing feature fusion enhancement processing on the medium and large size features after performing preliminary fusion enhancement processing on the feature data;

[0225] The road surface defects in the target image are detected according to the enhanced features.

[0226] Clause 2. The method according to clause 1, wherein the improved feature extraction network comprises a plurality of feature output layers;

[0227] Before the improved feature extraction network in the YOLOv5 model is used to perform feature extraction processing on the target image to obtain extracted feature data, the method further includes:

[0228] Set the number of CSP modules in multiple feature output layers;

[0229] Randomly select the number parameters corresponding to each feature output layer within a preset value range, and determine the corresponding candidate feature extraction network after each random selection;

[0230] Determining the target coverage corresponding to each of the candidate feature extraction networks;

[0231] Determining a preset number of optimal target coverage rates from the target coverage rates corresponding to each of the candidate feature extraction networks;

[0232] The improved feature extraction network is selected from a preset number of candidate feature extraction networks corresponding to better target coverage.

[0233] Clause 3. The method according to clause 2, wherein the number parameter corresponding to each feature output layer is randomly selected within a preset value range, comprising:

[0234] The number parameters of the CSP modules in each feature output layer are split into length dimension and width dimension to obtain the length number parameters and width number parameters corresponding to each feature output layer;

[0235] For each feature output layer, the length number parameter and the width number parameter are combined to form a number parameter combination corresponding to each feature output layer;

[0236] The length number parameter and the width number parameter in the number parameter combination corresponding to each feature output layer are randomly selected within a preset value range.

[0237] Clause 4. The method according to clause 2, wherein determining the target coverage corresponding to each of the candidate feature extraction networks comprises:

[0238] Calculate the receptive field size of each feature output layer of each candidate feature extraction network;

[0239] The coverage rate of different pavement diseases is calculated based on the receptive field size corresponding to each candidate feature extraction network;

[0240] According to the coverage rate of different pavement diseases corresponding to each candidate feature extraction network, the target coverage rate corresponding to each candidate feature extraction network is determined.

[0241] Clause 5. The method according to clause 2, wherein the step of selecting the improved feature extraction network from a preset number of candidate feature extraction networks corresponding to better target coverage ratios comprises:

[0242] Determine the parameter amount and computation amount of the candidate feature extraction network corresponding to a preset number of optimal target coverage rates;

[0243] An improved feature extraction network is selected from a preset number of candidate feature extraction networks according to a preset number of optimal target coverages, corresponding parameter quantities and calculation quantities.

[0244] Clause 6. The method according to any one of clauses 1 to 5, wherein the improved feature extraction network comprises: a plurality of feature extraction CSP modules, each CSP module being connected to a channel attention mechanism ECAM module;

[0245] The target image is subjected to feature extraction processing using an improved feature extraction network in the YOLOv5 model to obtain extracted feature data, including:

[0246] The CSP module is used to extract the features of the target image to obtain initial feature data;

[0247] The ECAM module is used to screen important features of the initial feature data to obtain extracted feature data.

[0248] Clause 7. The method according to clause 6, wherein the step of extracting features of the target image using a CSP module to obtain initial feature data comprises:

[0249] A main branch of a CSP module is used to perform multiple convolution operations, and a secondary branch of the CSP module is used to perform residual connection;

[0250] The concatenation operator Concat is used to concatenate the main branch processing result with the sub-branch processing result to obtain initial feature data.

[0251] Clause 8. The method according to clause 6, wherein the step of using an ECAM module to screen important features of the initial feature data to obtain extracted feature data comprises:

[0252] The ECAM module is used to calculate the weight vector of the initial feature data channel dimension;

[0253] Filter important feature data in the initial feature data according to the weight vector of the channel dimension;

[0254] The important feature data is determined as the extracted feature data.

[0255] Clause 9. The method according to any one of clauses 1 to 5, wherein the improved feature enhancement network comprises a network structure combining FPN and PAN, and a medium- and large-scale feature enhancement network structure;

[0256] The improved feature enhancement network in the YOLOv5 model is used to perform feature enhancement processing on the feature data to obtain enhanced features, including:

[0257] A network structure combining FPN and PAN is used to perform preliminary feature fusion and enhancement processing on the feature data to obtain preliminary enhanced features;

[0258] Using the medium- and large-scale feature enhancement network structure to perform feature fusion enhancement processing on the medium- and large-scale features based on the preliminary enhanced features, so as to obtain medium- and large-scale enhanced features;

[0259] The enhanced features are determined based on the preliminary enhanced features and the medium and large scale enhanced features.

[0260] Clause 10. The method according to clause 9, wherein the medium- and large-scale feature enhancement network structure comprises: a network structure combining multiple FPNs and PANs;

[0261] The method of using the medium- and large-scale feature enhancement network structure to perform feature fusion enhancement processing on the medium- and large-scale features based on the preliminary enhanced features to obtain medium- and large-scale enhanced features includes:

[0262] Inputting the medium-scale feature data and the large-scale feature data into the medium- and large-scale feature enhancement network structure;

[0263] The first FPN and PAN combined network structure is used to perform feature fusion and enhancement processing on the medium-scale feature data and the large-scale feature data to obtain fusion and enhanced features;

[0264] At least one FPN and PAN combined network structure in the middle position is used to perform feature fusion and enhancement processing on the previous fusion and enhancement features in turn, and the last FPN and PAN combined network structure outputs medium and large scale enhanced features.

[0265] Clause 11. A road surface quality detection device comprising:

[0266] An acquisition module, used for acquiring a target image to be identified, wherein the target image includes a road surface area;

[0267] A feature extraction module, used for performing feature extraction processing on the target image using an improved feature extraction network in a YOLOv5 model to obtain extracted feature data; the improved feature extraction network has a receptive field size suitable for pavement disease detection;

[0268] A feature enhancement module is used to perform feature enhancement processing on the feature data using the improved feature enhancement network in the YOLOv5 model to obtain enhanced features; the enhanced features are the features obtained by performing feature fusion enhancement on the medium and large size features after performing preliminary fusion enhancement processing on the feature data;

[0269] A detection module is used to detect road surface defects in the target image based on the enhanced features.

[0270] Clause 12. A road surface quality detection device, comprising: a processor and a memory; wherein:

[0271] The memory is used to store program code;

[0272] The processor is used to call the program code stored in the memory to execute the method described in any one of clauses 1-10.

[0273] Clause 13. An artificial intelligence chip, comprising a road surface quality detection device as described in Clause 11 or 12.

[0274] Clause 14. A board comprising: a storage device, an interface device, a control device and an artificial intelligence chip as described in Clause 13.

[0275] Clause 15. An electronic device comprising the artificial intelligence chip as described in Clause 13.

[0276] Clause 16. A computer-readable storage medium storing instructions, which, when executed on a computer, causes the computer to execute the method according to any one of clauses 1 to 10.

[0277] Clause 17. A computer program product comprising a computer program, which, when executed by a processor, implements the method as described in any one of clauses 1 to 10.

[0278] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily required by the present application.

[0279] It should be further noted that, although the various steps in the flowchart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowchart may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0280] It should be understood that the above-mentioned device embodiments are only illustrative, and the device of the present application can also be implemented in other ways. For example, the division of units / modules in the above-mentioned embodiments is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units, modules or components can be combined, or can be integrated into another system, or some features can be ignored or not executed.

[0281] In addition, unless otherwise specified, each functional unit / module in each embodiment of the present application may be integrated into one unit / module, each unit / module may exist physically separately, or two or more units / modules may be integrated together. The above-mentioned integrated unit / module may be implemented in the form of hardware or in the form of a software program module.

[0282] If the integrated unit / module is implemented in the form of hardware, the hardware can be a digital circuit, an analog circuit, etc. The physical implementation of the hardware structure includes but is not limited to transistors, memristors, etc. Unless otherwise specified, the artificial intelligence processor can be any appropriate hardware processor, such as CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the storage unit can be any appropriate magnetic storage medium or magneto-optical storage medium, such as resistive random access memory RRAM (Resistive Random Access Memory), dynamic random access memory DRAM (Dynamic Random Access Memory), static random access memory SRAM (Static Random-Access Memory), enhanced dynamic random access memory EDRAM (Enhanced Dynamic Random Access Memory), high-bandwidth memory HBM (High-Bandwidth Memory), hybrid memory cube HMC (Hybrid Memory Cube), etc.

[0283] If the integrated unit / module is implemented in the form of a software program module and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or all or part of the technical solution, can be embodied in the form of a software product, which is stored in a memory and includes several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the various embodiments of the present application. The aforementioned memory includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk and other media that can store program codes.

[0284] In the above embodiments, the description of each embodiment has its own emphasis. For the part not described in detail in a certain embodiment, please refer to the relevant description of other embodiments. The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A road quality detection method, It is characterized in that include: Acquire a target image to be identified, wherein the target image includes a road surface area; The target image is subjected to feature extraction processing by using an improved feature extraction network in a YOLOv5 model to obtain extracted feature data; the improved feature extraction network has a receptive field size suitable for pavement disease detection; the improved feature extraction network includes a plurality of feature output layers; The feature data is subjected to feature enhancement processing by using the improved feature enhancement network in the YOLOv5 model to obtain enhanced features; the enhanced features are the features obtained by performing feature fusion enhancement processing on the medium and large size features after performing preliminary fusion enhancement processing on the feature data; Detecting road surface defects in the target image according to the enhanced features; Before the improved feature extraction network in the YOLOv5 model is used to perform feature extraction processing on the target image to obtain extracted feature data, the method further includes: Set the number of CSP modules in multiple feature output layers; Randomly select the number parameters corresponding to each feature output layer within a preset value range, and determine the corresponding candidate feature extraction network after each random selection; Determining the target coverage corresponding to each of the candidate feature extraction networks; Determining a preset number of optimal target coverage rates from the target coverage rates corresponding to each of the candidate feature extraction networks; Determine the parameter amount and computation amount of the candidate feature extraction network corresponding to a preset number of optimal target coverage rates; If there is a candidate feature extraction network with the largest target coverage rate and the smallest parameter amount and calculation amount, the candidate feature extraction network with the largest target coverage rate and the smallest parameter amount and calculation amount is determined as the improved feature extraction network; If there is no candidate feature extraction network with the largest target coverage rate and the smallest parameter amount and calculation amount, the target coverage rate, parameter amount and calculation amount are normalized, and corresponding weights are set, and a weighted sum result is calculated by weighted summing the target coverage rate, parameter amount and calculation amount; the candidate feature extraction network with the largest weighted sum result is determined as the improved feature extraction network; The step of randomly selecting a value for a number parameter corresponding to each feature output layer within a preset value range includes: The number parameters of the CSP modules in each feature output layer are split into length dimension and width dimension to obtain the length number parameters and width number parameters corresponding to each feature output layer; For each feature output layer, the length number parameter and the width number parameter are combined to form a number parameter combination corresponding to each feature output layer; The length number parameter and the width number parameter in the number parameter combination corresponding to each feature output layer are randomly selected within a preset value range.

2. The method according to claim 1, It is characterized in that The determining of the target coverage rate corresponding to each of the candidate feature extraction networks includes: Calculate the receptive field size of each feature output layer of each candidate feature extraction network; The coverage rate of different pavement diseases is calculated based on the receptive field size corresponding to each candidate feature extraction network; According to the coverage rate of different pavement diseases corresponding to each candidate feature extraction network, the target coverage rate corresponding to each candidate feature extraction network is determined.

3. The method according to any one of claims 1 to 2, It is characterized in that The improved feature extraction network includes: a plurality of feature extraction CSP modules, each CSP module is connected with a channel attention mechanism ECAM module; The target image is subjected to feature extraction processing using an improved feature extraction network in the YOLOv5 model to obtain extracted feature data, including: The CSP module is used to extract the features of the target image to obtain initial feature data; The ECAM module is used to screen important features of the initial feature data to obtain extracted feature data.

4. The method according to claim 3, It is characterized in that The method of extracting the features of the target image using the CSP module to obtain initial feature data includes: The main branch of the CSP module is used to perform multiple convolution operations to obtain the main branch processing result, and the secondary branch of the CSP module is used to perform residual connection to obtain the secondary branch processing result; The concatenation operator Concat is used to concatenate the main branch processing result with the secondary branch processing result to obtain initial feature data.

5. The method according to claim 3, It is characterized in that The adopting of the ECAM module to perform important feature screening on the initial feature data to obtain extracted feature data includes: The ECAM module is used to calculate the weight vector of the initial feature data channel dimension; Filter important feature data in the initial feature data according to the weight vector of the channel dimension; The important feature data is determined as the extracted feature data.

6. The method according to any one of claims 1 to 2, It is characterized in that The improved feature enhancement network includes a network structure combining FPN and PAN, and a medium- and large-scale feature enhancement network structure; The improved feature enhancement network in the YOLOv5 model is used to perform feature enhancement processing on the feature data to obtain enhanced features, including: A network structure combining FPN and PAN is used to perform preliminary feature fusion and enhancement processing on the feature data to obtain preliminary enhanced features; Using the medium- and large-scale feature enhancement network structure to perform feature fusion enhancement processing on the medium- and large-scale features based on the preliminary enhanced features, so as to obtain medium- and large-scale enhanced features; The enhanced features are determined based on the preliminary enhanced features and the medium and large scale enhanced features.

7. The method according to claim 6, It is characterized in that The medium- and large-scale feature enhancement network structure includes: a network structure combining multiple FPNs and PANs; The method of using the medium- and large-scale feature enhancement network structure to perform feature fusion enhancement processing on the medium- and large-scale features based on the preliminary enhanced features to obtain medium- and large-scale enhanced features includes: Inputting the medium-scale feature data and the large-scale feature data into the medium- and large-scale feature enhancement network structure; The first FPN and PAN combined network structure is used to perform feature fusion and enhancement processing on the medium-scale feature data and the large-scale feature data to obtain fusion and enhanced features; At least one FPN and PAN combined network structure in the middle position is used to perform feature fusion and enhancement processing on the previous fusion and enhancement features in turn, and the last FPN and PAN combined network structure outputs medium and large scale enhanced features.

8. A road quality detection device, It is characterized in that include: An acquisition module, used for acquiring a target image to be identified, wherein the target image includes a road surface area; A feature extraction module, used for performing feature extraction processing on the target image using an improved feature extraction network in a YOLOv5 model to obtain extracted feature data; the improved feature extraction network has a receptive field size suitable for pavement disease detection; the improved feature extraction network includes a plurality of feature output layers; A feature enhancement module is used to perform feature enhancement processing on the feature data using the improved feature enhancement network in the YOLOv5 model to obtain enhanced features; the enhanced features are the features obtained by performing feature fusion enhancement on the medium and large size features after performing preliminary fusion enhancement processing on the feature data; A detection module, used for detecting road surface defects in the target image according to the enhanced features; The road surface quality detection device further includes: a first network determination module; The first network determination module is used to set the number parameter of CSP modules in multiple feature output layers; randomly select the number parameter corresponding to each feature output layer within a preset value range, and determine the corresponding candidate feature extraction network after each random selection; determine the target coverage rate corresponding to each candidate feature extraction network; determine a preset number of better target coverage rates from the target coverage rates corresponding to each candidate feature extraction network; determine the parameter amount and calculation amount of the candidate feature extraction network corresponding to the preset number of better target coverage rates; if there is a candidate feature extraction network with the largest target coverage rate and the smallest parameter amount and calculation amount, then the candidate feature extraction network with the largest target coverage rate and the smallest parameter amount and calculation amount is determined as the improved feature extraction network; if there is no candidate feature extraction network with the largest target coverage rate and the smallest parameter amount and calculation amount, then the target coverage rate, parameter amount and calculation amount are normalized, and the corresponding weights are set, and the weighted summation result is calculated by weighted summation of the target coverage rate, parameter amount and calculation amount; the candidate feature extraction network with the largest weighted summation result is determined as the improved feature extraction network; When the first network determination module randomly selects a value for the number parameter corresponding to each feature output layer within a preset value range, it is specifically used to: The number parameters of the CSP modules in each feature output layer are split into length dimension and width dimension to obtain the length number parameters and width number parameters corresponding to each feature output layer; For each feature output layer, the length number parameter and the width number parameter are combined to form a number parameter combination corresponding to each feature output layer; The length number parameter and the width number parameter in the number parameter combination corresponding to each feature output layer are randomly selected within a preset value range.

9. A road quality detection device, include: processor and memory; wherein, The memory is used to store program code; The processor is used to call the program code stored in the memory to execute the method according to any one of claims 1 to 7.

10. An artificial intelligence chip, It is characterized in that The artificial intelligence chip includes the road quality detection device as described in claim 8 or 9.

11. A board, It is characterized in that The board includes: a storage device, an interface device and a control device and the artificial intelligence chip as claimed in claim 10.

12. An electronic device, It is characterized in that The electronic device comprises the artificial intelligence chip as claimed in claim 10.

13. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the computer executes the method according to any one of claims 1 to 7.

14. A computer program product comprising a computer program, It is characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.