Road surface defect detection method and device

Through a deep learning network optimized by a maximum eight-direction defect detection filter and an edge information, the problem of distinguishing defects and background differences in road surface defect detection is solved, and the detection accuracy and efficiency are improved, especially showing superiority under different lighting conditions.

CN118674726BActive Publication Date: 2024-11-12SUZHOU CITY UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411172172.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-11-12
Estimated Expiration
2044-08-26

AI Technical Summary

Technical Problem

In the detection of road surface defects, it is difficult to effectively distinguish the differences between road defects and backgrounds. Especially under different lighting conditions, the detection accuracy is low, and the detection performance of irregular shape defects is poor, resulting in insufficient detection efficiency and accuracy. .

Method used

The maximum eight-direction defect detection filter is used to extract feature information from different directions, and the encoding-decoding deep learning network combines edge information to optimize the loss function, enhance defect characteristics, and distinguish road defects and backgrounds.

Benefits of technology

It improves the accuracy and efficiency of road surface defect detection, and can more accurately identify different types of defects under various lighting conditions, significantly improving detection accuracy and efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118674726B_ABST
    Figure CN118674726B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of surface defect recognition, and in particular to a road surface defect detection method and device, comprising: inputting an original image into a maximum eight-direction defect detection filter, wherein the maximum eight-direction defect detection filter comprises eight convolution kernels increasing at 45° angles, each convolution kernel is convolved with the original image to extract feature information from different directions; connecting all convolution results into a three-dimensional eight-channel data group, and taking the maximum value of pixels along the channel direction to obtain a feature-enhanced image; inputting the feature-enhanced image into a trained road surface defect detection model, and outputting a road surface defect recognition image. The present invention effectively distinguishes the difference between road defects and background, making different types of road surface defects easier to identify under various lighting conditions, thereby improving the accuracy and efficiency of defect detection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of surface defect recognition, and in particular to a road surface defect detection method and device. Background Art

[0002] Road surface defect detection is an indispensable procedure to ensure safe highway transportation. Traditional manual road surface defect detection is usually performed by trained manual inspectors, who need to check the road surface section by section for defects, which consumes a lot of manpower and time. In long-distance road inspections, manual road surface defect detection is extremely inefficient and difficult to meet actual needs. In addition, manual inspectors will inevitably become fatigued after working for a long time, resulting in inaccurate detection results. In addition, differences in subjective judgment between different manual inspectors will also affect the consistency of defect detection results. Moreover, it is difficult for manual inspectors to conduct inspections under severe weather conditions or in dangerous areas, which limits the continuity and coverage of inspections.

[0003] In order to solve the limitations of manual road surface defect detection, machine vision methods based on feature extraction have been applied to road surface defect detection, which can be mainly divided into three categories: texture feature-based methods, color feature-based methods, and shape feature-based methods. Although machine vision methods based on feature extraction have achieved certain success in surface defect detection, they still have obvious shortcomings and limitations in practical applications. Specifically, these methods have the following problems in terms of changes in detection accuracy under different ambient lighting conditions and adaptability to different types of defects: (1) The detection accuracy of the texture feature-based method is high under conditions of sufficient light on sunny days, but the detection accuracy drops sharply when the same section of road is detected on cloudy days or at night, indicating that changes in lighting conditions have a significant impact on the texture feature-based method; (2) For large-area, regular-shaped cracks and potholes on urban roads, the shape feature-based method can detect and locate them well, but for small, irregular-shaped cracks, there are problems with insufficient detection accuracy and high missed detection rates, indicating that the shape feature-based method performs poorly when dealing with irregular defects; (3) The texture feature-based method has good detection results under dry road conditions, but its detection accuracy drops significantly when the road is wet or covered with dust, indicating that environmental factors have a greater impact on the detection results; (4) Designing different feature extraction algorithms for different types of defects and performing complex parameter adjustments and optimizations will not only increase the cost of development and maintenance, but also easily affect the defect detection effect due to improper parameter settings.

[0004] In recent years, deep learning-based methods, especially convolutional neural networks (CNNs), have achieved remarkable results in many visual inspection projects. Deep learning has broken through the limitations of road surface defect detection, has a qualitative improvement in efficiency compared to traditional detection methods, and has also achieved significant improvements in detection performance. CNN-based surface defect detection can be roughly divided into two architectures: object detection-based networks and pixel segmentation-based networks. Pixel segmentation-based detection methods can not only detect defects, but also locate defects, so they have become the mainstream method. However, referring to Figure 1 As shown in Figure 1, the complexity of road surface defects poses a great challenge to computer vision-based detection. For low-contrast road defect images to be processed, the main difficulty is how to extract effective features to distinguish the difference between the target and the background. Specifically, (1) due to the uncertainty caused by the material itself and humidity and lighting, the contrast between defects and background in road images is low; (2) some road defects are small in size and cannot be detected in industrial images; (3) the shape and size of road defects are irregular.

[0005] Therefore, how to extract effective features to distinguish the difference between road defects and background and improve the efficiency and accuracy of road surface defect recognition methods is a technical problem that needs to be solved at present. Summary of the invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the problem in the prior art of how to extract effective features to distinguish the difference between road defects and background.

[0007] In order to solve the above technical problems, the present invention provides a road surface defect detection method, comprising:

[0008] The original image is input into a maximum eight-direction defect detection filter, and a feature-enhanced image is output; the maximum eight-direction defect detection filter includes eight convolution kernels increasing at 45° angles, each of which is convolved with the original image to extract feature information from different directions; all convolution results are connected into a three-dimensional eight-channel data group, and the maximum value of pixels along the channel direction is taken to obtain a feature-enhanced image;

[0009] The feature enhanced image is input into the trained road surface defect detection model and the road surface defect recognition image is output.

[0010] Preferably, the formula of the maximum eight-directional defect detection filter is expressed as:

[0011] ;

[0012] ;

[0013] in, represents the original image, represents the nth convolution kernel, is the convolution operator, is the convolution result of the original image and the nth convolution kernel, Enhance images for features.

[0014] Preferably, the eight convolution kernels increasing at an angle of 45° are a 5×5 convolution kernel.

[0015] Preferably, the eight convolution kernels increasing at 45° angles are 5×5 convolution kernels and specifically include:

[0016] ;

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] ;

[0023] ;

[0024] in, represents the nth convolution kernel, .

[0025] Preferably, the road surface defect detection model is an encoding-decoding deep learning network.

[0026] Preferably, during the training process of the road surface defect detection model, the loss function is:

[0027] ;

[0028] in, is the loss function of the road surface defect detection model, is the edge information function, Annotate images for road surface defects, A road surface defect recognition image output by a road surface defect detection model; is the weight matrix, is a hyperparameter used to determine the weight of the pixel loss term, represents a weight matrix whose elements are all 1.

[0029] Preferably, the edge information function The formula is:

[0030] ;

[0031] in, Represents the weight matrix In Location The weight of Represents the road surface defect recognition image output by the road surface defect detection model at position The pixel value of Indicates that the road surface defect annotation image is at position The pixel value of , k is the channel index, i and j represent the horizontal and vertical coordinates of the pixel in the image respectively.

[0032] Preferably, the edge area is emphasized when training the road surface defect detection model, and the loss function of the road surface defect detection model is optimized, and the optimized loss function is obtained as follows:

[0033] ;

[0034] in, is the loss function of the optimized road surface defect detection model, is a hyperparameter, is the edge information weight obtained using the local variance method.

[0035] Preferably, the edge information weight obtained by the local variance method is The formula is:

[0036] ;

[0037] ;

[0038] in, Indicates that the image is labeled Middle position The edge information weight of is, k is the channel index, i and j represent the horizontal and vertical coordinates of the pixel in the image respectively, Indicates surrounding The sliding window represents the number of pixels in the sliding window, and Respectively represent the step size of calculating the variance of each pixel in the sliding window in the horizontal and vertical directions, Represents annotated image In Location The pixel value of .

[0039] The present invention also provides a road surface defect detection device, comprising:

[0040] A feature enhancement module is used to input the original image into a maximum eight-directional defect detection filter and output a feature enhanced image; the maximum eight-directional defect detection filter includes 8 convolution kernels increasing at an angle of 45°, each convolution kernel is convolved with the original image to extract feature information from different directions; all convolution results are connected into a three-dimensional eight-channel data group, and the maximum value of pixels along the channel direction is taken to obtain a feature enhanced image;

[0041] The defect recognition module is used to input the feature enhanced image into the trained road surface defect detection model and output the road surface defect recognition image.

[0042] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0043] A road surface defect detection method described in the present invention adopts a maximum eight-directional defect detection filter to extract defect features in an original image from different directions, suppresses background information of the original image by taking the maximum pixel value in the channel direction, thereby enhancing the defect features and effectively distinguishing the difference between road defects and the background, making different types of road surface defects easier to identify under various lighting conditions, thereby improving the accuracy and efficiency of defect detection.

[0044] Furthermore, when training a road surface defect detection model, the present invention optimizes the loss function based on the edge information of the defects, combines the low-frequency background information and the high-frequency edge information, and selectively extracts the edge information, so that the difference between the road defects and the background is further enhanced, thereby further improving the accuracy of defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0046] Figure 1 is an example diagram of road surface defects;

[0047] Figure 2 is a flow chart of a road surface defect detection method of the present invention;

[0048] Figure 3 is a schematic diagram of a maximum eight-directional defect detection filter;

[0049] Figure 4 is a schematic diagram of annotated images used in the present invention, wherein Figure 4 (a) is the original image of the first road surface defect. Figure 4 (b) in the figure is the labeled image of the first road surface defect. Figure 4(c) in the figure is the original image of the second road surface defect. Figure 4 (d) in the figure is the annotated image of the second road surface defect;

[0050] Figure 5 It is a schematic diagram of loss function calculation of the present invention;

[0051] Figure 6 is a comparison diagram of the original image and the road surface defect recognition image output by the present invention, wherein Figure 6 (a) is the original image of the third road surface defect. Figure 6 (b) is the recognition image of the third road surface defect. Figure 6 (c) in the figure is the original image of the fourth road surface defect. Figure 6 (d) in the figure is a recognition image of the fourth road surface defect. DETAILED DESCRIPTION

[0052] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0053] After in-depth research and analysis in the early stage, the applicant found that road surface defects in the image will destroy the smoothness and consistency of the entire image. Even the smallest defects will affect the consistency of the image. Therefore, in response to this discovery, the present invention proposes a road surface defect detection method. The flowchart is shown in FIG. Figure 2 As shown, including:

[0054] The original image is input into a maximum eight-direction defect detection filter, and a feature-enhanced image is output; the maximum eight-direction defect detection filter includes eight convolution kernels increasing at 45° angles, each of which is convolved with the original image to extract feature information from different directions; all convolution results are connected into a three-dimensional eight-channel data group, and the maximum value of pixels along the channel direction is taken to obtain a feature-enhanced image;

[0055] The feature enhanced image is input into the trained road surface defect detection model and the road surface defect recognition image is output.

[0056] The schematic diagram of the maximum eight-directional defect detection filter is shown in FIG. Figure 3 As shown, the formula is expressed as:

[0057] ;

[0058] ;

[0059] in, represents the original image, represents the nth convolution kernel, is the convolution operator, is the convolution result of the original image and the nth convolution kernel, Enhance images for features.

[0060] Preferably, the eight convolution kernels increasing at 45° angles in the maximum eight-directional defect detection filter are 5×5 convolution kernels, specifically including:

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] ;

[0067] ;

[0068] ;

[0069] in, represents the nth convolution kernel, .

[0070] Since road surface defects are outliers compared to the smooth background area in the original image, the defects are enhanced through convolution in eight directions after passing through the maximum eight-direction defect detection filter, and finally the maximum pixel value is taken in the channel direction, which can further suppress the background area and enhance the defective part in the image.

[0071] Preferably, the road surface defect detection model is an encoding-decoding deep learning network.

[0072] During the training process of the road surface defect detection model, the distance loss is used as the training loss or a part of the training loss. The distance loss is to calculate the dissimilarity between two images. In the road surface defect detection model involved in the present invention, it is the difference between the annotated image and the recognized image, which is usually calculated in the pixel and feature space. The annotated image used in the present invention refers to Figure 4 As shown, Figure 4 (a) is the original image of the first road surface defect. Figure 4 (b) in the figure is the labeled image of the first road surface defect. Figure 4 (c) in the figure is the original image of the second road surface defect. Figure 4(d) in the figure is the annotated image of the second road surface defect. The goal of the road surface defect detection model is to minimize the distance loss, so as to push the recognition result to be close to the corresponding true value image. The schematic diagram of the loss function calculation is shown in Figure 5 shown.

[0073] Through the analysis of road crack images, we can see that road cracks have clear boundaries in the image, that is, edge information, while the road part without crack defects shows a low-frequency feature with gentle fluctuations. Therefore, edge information is the key information that can be used as crack identification.

[0074] Since the edge information in the road image is extremely critical to the accuracy of recognition, in order to retain the edge information in the road surface defect recognition image as much as possible, the weight of the high-frequency area, that is, the edge area, should be greater than the low-frequency area, that is, the background area, in the training model. The road surface defect detection model of the present invention adopts a loss function based on edge information during training, and the formula is:

[0075] ;

[0076] in, is the loss function of the road surface defect detection model, is the edge information function, Annotate images for road surface defects, A road surface defect recognition image output by a road surface defect detection model; is the weight matrix, is a hyperparameter used to determine the weight of the pixel loss term, represents a weight matrix whose elements are all 1.

[0077] Edge Information Function The formula is:

[0078] ;

[0079] in, Represents the weight matrix In Location The weight of Represents the road surface defect recognition image output by the road surface defect detection model at position The pixel value of Indicates that the road surface defect annotation image is at position The pixel value of .

[0080] Furthermore, in order to emphasize the edge area during training, this embodiment selects edge information as the main representative of spatial features, uses the local variance method to detect and extract key edge information, and optimizes the loss function of the road surface defect detection model. The optimized loss function based on edge information is:

[0081] ;

[0082] in, is the loss function of the optimized road surface defect detection model, is a hyperparameter, which is the coefficient of the pixel loss used in the training loss; is the edge information weight obtained using the local variance method.

[0083] Edge information weight obtained using local variance method The formula is:

[0084] ;

[0085] ;

[0086] in, Indicates that the image is labeled Middle position The edge information weight of is, k is the channel index, i and j represent the horizontal and vertical coordinates of the pixel in the image respectively, Indicates surrounding The sliding window represents the number of pixels in the sliding window, and Respectively represent the step size of calculating the variance of each pixel in the sliding window in the horizontal and vertical directions, Represents annotated image In Location The pixel value of .

[0087] The optimized loss function will treat edge pixels as more important pixels during training and invest more energy in recovering them. In other words, during the reverse pass, the trainable parameters will be "consciously" updated in the direction of more precise edges that can be super-resolved, and the parameters of those edge areas will be given greater weight updates.

[0088] Figure 6 It is a comparison diagram of the original image and the road surface defect recognition image output by the present invention, wherein Figure 6 (a) is the original image of the third road surface defect. Figure 6 (b) is the recognition image of the third road surface defect. Figure 6(c) in the figure is the original image of the fourth road surface defect. Figure 6 (d) in the figure is the identification image of the fourth road surface defect.

[0089] In summary, a road surface defect detection method described in the present invention adopts a maximum eight-directional defect detection filter to extract defect features in the original image from different directions, and suppresses the background information of the original image by taking the maximum pixel value in the channel direction, so that the defect features are enhanced, and the difference between road defects and the background is effectively distinguished, so that different types of road surface defects are easier to identify under various lighting conditions, thereby improving the accuracy and efficiency of defect detection.

[0090] Furthermore, when training a road surface defect detection model, the present invention optimizes the loss function based on the edge information of the defects, combines the low-frequency background information and the high-frequency edge information, and selectively extracts the edge information, so that the difference between the road defects and the background is further enhanced, thereby further improving the accuracy of defect detection.

[0091] In order to verify the superiority of the method proposed in the present invention, this embodiment has conducted a large number of experiments on multiple public surface defect datasets, including DCD dataset, CSDD dataset and FSDD dataset, and compared the detection effects under different types of defects and different environmental conditions. The experimental results are shown in the following table:

[0092] Table 1. Comparison of detection results of different data sets

[0093] Dataset Canny edge detection algorithm The present invention Improved efficiency DCD 72% 90% 18% CSDD 75% 88% 23% FSDD 70% 92% 22%

[0094] Table 2. Comparison of detection results of different defect types

[0095] Defect Type Canny edge detection algorithm The present invention Improved efficiency crack 78% 93% 18% Potholes 68% 90% 22% Flaking 70% 91% 21% other 65% 89% 24%

[0096] Table 3. Comparison of test results under different environmental conditions

[0097] Environmental conditions Canny edge detection algorithm The present invention Improved efficiency Adequate lighting 85% 98% 10% Insufficient light 50% 85% 35% No light at night 30% 80% 50%

[0098] It can be seen from the experimental data that the detection accuracy of the road surface defect detection method proposed by the present invention is significantly higher than that of the traditional Canny edge detection algorithm on all data sets, and the detection efficiency has been greatly improved. In addition, the maximum eight-directional defect detection filter and the loss function based on edge information proposed by the present invention show significant superiority in different data sets and different lighting conditions, and can effectively improve the detection accuracy and efficiency.

[0099] Based on the above-mentioned road surface defect detection method, this embodiment further provides a road surface defect detection device, including:

[0100] A feature enhancement module is used to input the original image into a maximum eight-directional defect detection filter and output a feature enhanced image; the maximum eight-directional defect detection filter includes 8 convolution kernels increasing at an angle of 45°, each convolution kernel is convolved with the original image to extract feature information from different directions; all convolution results are connected into a three-dimensional eight-channel data group, and the maximum value of pixels along the channel direction is taken to obtain a feature enhanced image;

[0101] The defect recognition module is used to input the feature enhanced image into the trained road surface defect detection model and output the road surface defect recognition image.

[0102] Preferably, the formula of the maximum eight-directional defect detection filter in the feature enhancement module is expressed as:

[0103] ;

[0104] ;

[0105] in, represents the original image, represents the nth convolution kernel, is the convolution operator, is the convolution result of the original image and the nth convolution kernel, Enhance images for features.

[0106] Preferably, the eight convolution kernels increasing at 45° angles in the maximum eight-directional defect detection filter are 5×5 convolution kernels.

[0107] Preferably, in the defect recognition module, during the training process of the road surface defect detection model, the loss function is:

[0108] ;

[0109] in, is the loss function of the road surface defect detection model, is the edge information function, Annotate images for road surface defects, A road surface defect recognition image output by a road surface defect detection model; is the weight matrix, is a hyperparameter used to determine the weight of the pixel loss term, represents a weight matrix whose elements are all 1.

[0110] Preferably, the edge information function The formula is:

[0111] ;

[0112] in, Represents the weight matrix In Location The weight of Represents the road surface defect recognition image output by the road surface defect detection model at position The pixel value of Indicates that the road surface defect annotation image is at position The pixel value of .

[0113] Preferably, in the defect recognition module, the edge area can be further emphasized when training the road surface defect detection model, and the loss function of the road surface defect detection model is optimized, and the optimized loss function is:

[0114] ;

[0115] in, is the loss function of the optimized road surface defect detection model, is a hyperparameter, is the edge information weight obtained using the local variance method.

[0116] Preferably, the edge information weight obtained by the local variance method is The formula is:

[0117] ;

[0118] ;

[0119] in, Indicates that the image is labeled Middle position The edge information weight of is, k is the channel index, i and j represent the horizontal and vertical coordinates of the pixel in the image respectively, Indicates surrounding The sliding window represents the number of pixels in the sliding window, and Respectively represent the step size of calculating the variance of each pixel in the sliding window in the horizontal and vertical directions, Represents annotated image In Location The pixel value of .

[0120] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0121] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0122] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0123] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0124] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A road surface defect detection method, characterized in that: include: The original image is input into the maximum eight-directional defect detection filter, and the feature enhanced image is output; The maximum eight-direction defect detection filter includes eight convolution kernels increasing at 45° angles, each of which is convolved with the original image to extract feature information from different directions; all convolution results are connected into a three-dimensional eight-channel data set, and the maximum value of pixels is taken along the channel direction to obtain a feature enhanced image; The eight convolution kernels increasing at 45° angles are 5×5 convolution kernels, including: in, represents the nth convolution kernel, n = 1, 2, ..., 8; The feature enhanced image is input into the trained road surface defect detection model and the road surface defect recognition image is output.

2. A road surface defect detection method according to claim 1, characterized in that: The formula of the maximum eight-directional defect detection filter is expressed as: Among them, I represents the original image, represents the nth convolution kernel, * is the convolution operator, is the convolution result of the original image and the nth convolution kernel, I MEDS Enhance images for features.

3. A road surface defect detection method according to claim 1, characterized in that: The road surface defect detection model is an encoding-decoding deep learning network.

4. A road surface defect detection method according to claim 1, characterized in that: During the training process of the road surface defect detection model, the loss function is: in, is the loss function of the road surface defect detection model, l w is the edge information function, Gr is the road surface defect annotation image, Pr is the road surface defect recognition image output by the road surface defect detection model; α1 is the weight matrix, α is a hyperparameter used to determine the weight of the pixel loss term, and 1 represents a weight matrix whose elements are all 1.

5. A road surface defect detection method according to claim 4, characterized in that: Edge information function l w The formula is: Among them, h k,i,j represents the weight of the weight matrix α1 at position (k,i,j), Pr k,i,j Represents the pixel value of the road surface defect recognition image at position (k,i,j) output by the road surface defect detection model, Gr k,i,j Represents the pixel value of the road surface defect annotation image at position (k,i,j), where k is the channel index, and i and j represent the horizontal and vertical coordinates of the pixel in the image, respectively.

6. A road surface defect detection method according to claim 5, characterized in that: When training the road surface defect detection model, the edge area is emphasized, and the loss function of the road surface defect detection model is optimized. The optimized loss function is: in, is the loss function of the optimized road surface defect detection model, β is a hyperparameter, and μ(Gr) is the edge information weight obtained using the local variance method.

7. A road surface defect detection method according to claim 6, characterized in that: The formula of edge information weight μ(Gr) obtained by local variance method is: Among them, μ k,i,j (Gr) represents the edge information weight of position (k,i,j) in the labeled image Gr, k is the channel index, i and j represent the horizontal and vertical coordinates of the pixel in the image, respectively. Γ k,i,j represents the sliding window around Gr(k,i,j), |Γ k,i,j | represents the number of pixels in the sliding window, l1 and l2 represent the step size for calculating the variance of each pixel in the sliding window in the horizontal and vertical directions, respectively, and Gr(k,i+l1,j+l2) represents the pixel value of the labeled image Gr at position (k,i+l1,j+l2).

8. A road surface defect detection device, characterized in that: include: A feature enhancement module, used to input the original image into a maximum eight-directional defect detection filter and output a feature enhanced image; The maximum eight-direction defect detection filter includes eight convolution kernels increasing at 45° angles, each of which is convolved with the original image to extract feature information from different directions; all convolution results are connected into a three-dimensional eight-channel data set, and the maximum value of pixels is taken along the channel direction to obtain a feature enhanced image; The eight convolution kernels increasing at 45° angles are 5×5 convolution kernels, including: in, represents the nth convolution kernel, n = 1, 2, ..., 8; The defect recognition module is used to input the feature enhanced image into the trained road surface defect detection model and output the road surface defect recognition image.

Citation Information

Patent Citations

  • Fan blade defect detection method and system based on machine vision

    CN112098416A

  • Appearance defect detection method and system

    CN112950618A

  • Silicon wafer defect classification method based on combined convolution

    CN117132807A

  • Method and device for detecting stripe defective of picture and display device

    JP2005165387A