Road surface defect detection system and method based on deep learning

Abstract

The invention discloses a pavement defect detection system and method based on deep learning, and the system comprises an image collection module which is configured in an unmanned aerial vehicle and is used for collecting an image of a to-be-detected pavement when the unmanned aerial vehicle flies along a preset path to obtain a pavement image; the image transmission module is connected with the image acquisition module; the image processing module is connected with the image transmission module, and the image processing module is used for receiving the road surface image transmitted by the image transmission module and sequentially carrying out preprocessing, feature extraction and defect recognition on the road surface image based on a deep learning network so as to generate a defect recognition result; and the result output module is connected with the image processing module, and the result output module is used for generating a pavement defect detection report according to the defect recognition result. According to the invention, the detection efficiency and the detection precision of pavement defect detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of road surface defect detection, and more specifically to a road surface defect detection system and method based on deep learning. Background Art

[0002] In the modern transportation system, the quality of highway road conditions is directly related to driving safety. Existing road surface defect detection methods mainly include manual inspection, fixed cameras, and road surface inspection based on traditional image processing technology. The problems are as follows: (1) Manual inspection is time-consuming and costly, making it difficult to cover large road sections, and the detection quality is easily affected by human factors; (2) Fixed cameras lack flexibility due to their fixed monitoring position, and image quality degrades under adverse weather conditions, affecting detection accuracy; (3) Traditional image processing technology has poor adaptability in complex road environments, has a high false detection rate, and is difficult to accurately identify various types of defects such as cracks and potholes. Summary of the Invention

[0003] The purpose of the present invention is to provide a road surface defect detection system and method based on deep learning to solve the problem of low detection efficiency and detection accuracy of existing road surface defect detection methods.

[0004] In order to achieve the above-mentioned objectives, in a first aspect, the present invention provides a road surface defect detection system based on deep learning, comprising: an image acquisition module, which is configured in a drone, and the image acquisition module is used to acquire an image of the road surface to be detected when the drone flies along a preset path to obtain a road surface image; an image transmission module, which is connected to the image acquisition module; an image processing module, which is connected to the image transmission module, and the image processing module is used to receive the road surface image transmitted by the image transmission module, and perform preprocessing, feature extraction and defect recognition on the road surface image in sequence based on a deep learning network to generate a defect recognition result; a result output module, which is connected to the image processing module, and the result output module is used to generate a road surface defect detection report according to the defect recognition result.

[0005] Its further technical solution is: the image processing module includes: an image preprocessing submodule, used to preprocess the road surface image to obtain a target road surface image; a feature extraction submodule, used to extract features of the target road surface image through a deep learning network to obtain road surface features; a defect recognition submodule, used to classify the road surface features through a trained defect recognition model to obtain the defect recognition results.

[0006] Its further technical solution is: the image preprocessing submodule includes: a filtering submodule, which is used to use Gaussian filtering or median filtering to eliminate noise in the road surface image to obtain a filtered road surface image; and a clarity adjustment submodule, which is used to adjust the clarity of the filtered road surface image through a histogram equalization method or an adaptive threshold adjustment technology to obtain the target road surface image.

[0007] Its further technical solution is: the drone is also equipped with an auxiliary acquisition device, and the image processing module also includes: a defogging submodule, which is used to restore the clarity of the road image collected by the auxiliary acquisition device through a dark channel prior algorithm under foggy conditions to obtain a clear road image, so that the image preprocessing submodule preprocesses the clear road image to obtain the target road image.

[0008] Its further technical solution is: the image processing module also includes: a brightness adjustment submodule, which is used to improve the brightness of the road surface image collected by the auxiliary acquisition device under low light conditions through a gamma correction method or a CLAHE algorithm to obtain a brightness road surface image, so that the image preprocessing submodule preprocesses the brightness road surface image to obtain the target road surface image.

[0009] Its further technical solution is: the road surface defect detection system also includes: a first exception handling module, used to control the drone to return and save the collected road surface image when the drone is in an abnormal state; a second exception handling module, used to cache the road surface image when the road surface image transmission is interrupted, and continue to transmit the road surface image after the connection is restored; a third exception handling module, used to issue a prompt for manual review when the defect recognition model classification is wrong.

[0010] Its further technical solution is: the road surface defect detection system also includes: a distributed module, which is used to distribute the road surface images collected by the image acquisition module in the drone in different sections to different threads when detecting large-scale road surfaces, so that the threads call the image processing module to perform parallel processing on the road surface images in different sections, obtain multiple defect identification sub-results, and generate the road surface defect detection report through the result output module based on the multiple defect identification sub-results.

[0011] In order to achieve the above-mentioned purpose, in a second aspect, the present invention also provides a pavement defect detection method based on deep learning, which is applied to the pavement defect detection system based on deep learning described in the first aspect above, and the method includes: through an image acquisition module configured in a drone, when the drone flies along a preset path, collecting an image of the road surface to be detected to obtain a pavement image; connecting with the image acquisition module through an image transmission module; connecting with the image transmission module through an image processing module to receive the pavement image transmitted by the image transmission module, and preprocessing, feature extraction and defect recognition of the pavement image in sequence based on a deep learning network to generate a defect recognition result; connecting with the image processing module through a result output module to generate a pavement defect detection report according to the defect recognition result.

[0012] Its further technical solution is: the road surface image is preprocessed, feature extracted and defect identified in sequence based on the deep learning network to generate a defect identification result, including: preprocessing the road surface image to obtain a target road surface image; extracting features of the target road surface image through a deep learning network to obtain road surface features; and classifying the road surface features through a trained defect recognition model to obtain the defect identification result, wherein the defect recognition model is a model obtained by training and verifying a convolutional neural network using data sets under various conditions.

[0013] Its further technical solution is: the data sets under various conditions include sample road surface data sets collected on sunny days, rainy days, foggy days and at night, and the data sets under various conditions are used to train and verify the convolutional neural network to obtain the defect recognition model, including: using data enhancement technology to perform data enhancement on the sample road surface data sets collected on sunny days, rainy days, foggy days and at night to obtain an enhanced data set, and labeling the enhanced data set, and dividing the labeled enhanced data set into a training data set and a verification data set; inputting the training data set into the convolutional neural network for training until the confidence of the defect category output by the convolutional neural network exceeds the preset confidence; inputting the verification data set into the trained convolutional neural network for verification to output the verification defect category; calculating the recognition defect accuracy rate based on the verification defect category and the defect category in the verification data set; if the recognition defect accuracy rate is greater than the preset accuracy rate, the trained convolutional neural network is used as the defect recognition model.

[0014] An embodiment of the present invention provides a road surface defect detection system and method based on deep learning. First, an image acquisition module is configured in a drone to acquire an image of the road surface to be detected when the drone flies along a preset path to obtain a road surface image; then, an image processing module is used to preprocess, extract features, and identify defects on the road surface image transmitted by the image transmission module based on a deep learning network to generate a defect identification result; and a road surface defect detection report is generated through a result output module based on the defect identification result, thereby achieving efficient and automated detection of road surface defects, which not only improves the detection efficiency, but also improves the detection precision due to the introduction of the deep learning network, thereby improving the accuracy of the detection.

[0015] The present invention will become more apparent from the following description taken in conjunction with the accompanying drawings, which are used to illustrate embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A block diagram of a road surface defect detection system based on deep learning according to the present invention; Figure 2 This is a block diagram of an image processing module in a road surface defect detection system based on deep learning according to the present invention; Figure 3 This is a block diagram of an image processing submodule in a road surface defect detection system based on deep learning according to the present invention; Figure 4 is another block diagram of a road surface defect detection system based on deep learning according to the present invention; Figure 5 is another block diagram of a road surface defect detection system based on deep learning according to the present invention; Figure 6 This is a flowchart of a road surface defect detection method based on deep learning according to an embodiment of the present invention; Reference numerals: 10. Road defect detection system based on deep learning; 11. Image acquisition module; 12. Image transmission module; 13. Image processing module; 131. Image preprocessing submodule; 1311. Filtering submodule; 1312. Clarity adjustment submodule; 132. Feature extraction submodule; 133. Defect recognition submodule; 134. Defogging submodule; 135. Brightness adjustment submodule; 14. Result output module; 15. First exception handling module; 16. Second exception handling module; 17. Third exception handling module; 18. Distributed module. DETAILED DESCRIPTION

[0017] The following will be combined with the accompanying drawings of the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments. Similar component numbers in the drawings represent similar components. Obviously, the embodiments described below are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] Reference Figures 1 to 5 The deep learning-based pavement defect detection system 10 provided in an embodiment of the present invention includes an image acquisition module 11, an image transmission module 12, an image processing module 13 and a result output module 14, wherein the image acquisition module 11 is configured in a drone, and the image acquisition module 11 is used to acquire an image of the road surface to be detected when the drone flies along a preset path to obtain a pavement image; the image transmission module 12 is connected to the image acquisition module 11; the image processing module 13 is connected to the image transmission module 12, and the image processing module 13 is used to receive the pavement image transmitted by the image transmission module 12, and preprocess, extract features and identify defects on the pavement image in sequence based on a deep learning network to generate a defect identification result; the result output module 14 is connected to the image processing module 13, and the result output module 14 is used to generate a pavement defect detection report according to the defect identification result. It should be noted that, in this embodiment, the image acquisition module 11 is a high-definition camera, the UAV maintains a flight altitude of 10-20 meters, flies along a preset path at a speed of 5-10 meters per second, and the high-definition camera shoots at a fixed frequency (about 5 frames per second) to ensure full coverage of the road surface. The preset path ensures that the detection area is not missed, thereby achieving efficient road image acquisition; the data transmission module is a 4G network or a 5G network, and the road image is transmitted from the UAV to the server via the 4G network or the 5G network. The server is a cloud server or a ground station server, and the server is configured with an image processing module 13 and a result output module 14. It should also be noted that, in this embodiment, the image acquisition module 11 is first configured in the UAV to acquire an image of the road surface to be detected when the UAV flies along a preset path; then the image processing module 13 preprocesses, extracts features, and identifies defects on the road surface image transmitted by the image transmission module 12 based on the deep learning network to generate a defect recognition result, and a road surface defect detection report is generated through the result output module 14 according to the defect recognition result, thereby achieving efficient and automated detection of road surface defects, which not only improves the detection efficiency, but also improves the detection precision due to the introduction of the deep learning network, thereby improving the accuracy of the detection.

[0019] In some embodiments, such as the present embodiment, Figure 2As shown, the image processing module 13 includes an image preprocessing submodule 131, a feature extraction submodule 132, and a defect recognition submodule 133, wherein the image preprocessing submodule 131 is used to preprocess the road surface image to obtain a target road surface image; the feature extraction submodule 132 is used to extract features from the target road surface image through a deep learning network to obtain road surface features; and the defect recognition submodule 133 is used to classify the road surface features using a trained defect recognition model to obtain the defect recognition result. It should be noted that, in this embodiment, the deep learning network can be any one of CNN (convolutional neural network), ResNet (residual network), and EfficientNet (efficient network). It is understandable that ResNet and EfficientNet networks can improve feature extraction accuracy and efficiency. It should also be noted that, in this embodiment, the road surface features include image texture features and image edge features.

[0020] In some embodiments, such as the present embodiment, Figure 3 As shown, the image preprocessing submodule 131 includes a filtering submodule 1311 and a clarity adjustment submodule 1312. The filtering submodule 1311 is configured to eliminate noise from the road surface image using Gaussian filtering or median filtering to obtain a filtered road surface image. The clarity adjustment submodule 1312 is configured to adjust the clarity of the filtered road surface image using a histogram equalization method or an adaptive threshold adjustment technique to obtain the target road surface image. It should be noted that in this embodiment, adaptive threshold adjustment is a technique for dynamically calculating the threshold value of a local area of ​​an image, which can effectively address issues such as uneven illumination and low contrast, thereby improving the clarity of the filtered road surface image. Histogram equalization is an image processing technique that enhances contrast by adjusting the image pixel distribution. Its core concept is to convert the image's histogram (pixel value distribution) into a uniform distribution, thereby expanding the dynamic range of pixel values ​​and improving visual effects.

[0021] In some embodiments, such as the present embodiment, Figure 2As shown, the drone is also equipped with an auxiliary acquisition device. The image processing module 13 also includes a defogging module 134. This defogging module 134 is used to restore the clarity of the road surface image captured by the auxiliary acquisition device under foggy conditions using a dark channel prior algorithm to obtain a clear road surface image, enabling the image preprocessing submodule 131 to preprocess the clear road surface image to obtain the target road surface image. It should be noted that in this embodiment, the auxiliary acquisition device is an infrared camera or a lidar. By integrating an infrared camera or lidar into the drone to assist in acquiring road surface images, the robustness and accuracy of the road surface defect detection system are ensured. The dark channel prior algorithm was originally used for image defogging and was later expanded to low-light enhancement. Its core assumption is that in most non-sky areas, at least one color channel has pixel values ​​close to 0 (dark channel prior). Image clarity restoration is achieved by estimating atmospheric light and transmittance.

[0022] In some embodiments, such as the present embodiment, Figure 2 As shown, the image processing module 13 also includes a brightness adjustment submodule 135. This brightness adjustment submodule 135 is configured to, under low-light conditions, enhance the brightness of the road surface image captured by the auxiliary acquisition device using a gamma correction method or a CLAHE algorithm to obtain a luminance road surface image, enabling the image preprocessing submodule 131 to preprocess the luminance road surface image to obtain the target road surface image. It should be noted that in this embodiment, the gamma correction method adjusts the image brightness distribution through a power-law transformation; CLAHE (Contrast Limited Adaptive Histogram Equalization) is an improved form of histogram equalization. In this embodiment, the CLAHE algorithm divides the image into multiple tiles and performs histogram equalization on each tile. The clipLimit parameter limits the contrast enhancement amplitude to avoid noise amplification. The clipLimit parameter serves as the clipping threshold for controlling the subtile histogram.

[0023] In some embodiments, such as the present embodiment, Figure 4As shown, the pavement defect detection system further includes a first exception handling module 15, a second exception handling module 16, and a third exception handling module 17. The first exception handling module 15 is used to control the drone to return and save the collected pavement image when the drone is in an abnormal state; the second exception handling module 16 is used to cache the pavement image when the pavement image transmission is interrupted, and to continue transmitting the pavement image after the connection is restored; the third exception handling module 17 is used to issue a manual review prompt when the defect recognition model classification is incorrect. It should be noted that in this embodiment, the first exception handling module 15 and the second exception handling module 16 are both configured in the drone; the third exception handling module 17 is set in the server. The setting of the third exception handling module 17 can ensure the accuracy of pavement defect recognition.

[0024] In some embodiments, such as the present embodiment, Figure 5 As shown, the pavement defect detection system also includes a distributed module 18, which is used to distribute the pavement images collected by the image acquisition module 11 in the drone in different sections to different threads when inspecting large-scale road surfaces, so that the threads call the image processing module 13 to perform parallel processing on the pavement images in different sections, obtain multiple defect identification sub-results, and generate the pavement defect detection report through the result output module 14 based on the multiple defect identification sub-results. It should be noted that, in this embodiment, the pavement images collected by the image acquisition module 11 in the drone in different sections are distributed to different threads through the distributed module 18, so that the threads call the image processing module 13 to perform parallel processing on the pavement images in different sections, thereby improving the efficiency of pavement image processing. It should also be noted that, in this embodiment, the pavement defect detection report includes all section defect statistics and maintenance recommendations.

[0025] For ease of understanding, the following describes in detail the deep learning-based road surface defect detection system 10 from three examples: basic road surface defect detection, road surface defect detection in severe weather, and large-scale road network detection. In one embodiment (basic pavement defect detection), a quadrotor drone equipped with a 20-megapixel high-definition camera is used, flying at an altitude of 15 meters, at a speed of 8 meters per second, and capturing 5 frames per second. A ground station server is equipped with an NVIDIA RTX 3080 GPU running a deep learning network and defect recognition model. The pavement defect detection steps include: the drone flies along a preset path to capture pavement images; the images are transmitted in real time via a 4G network with a transmission delay of less than 1 second; the ground station server performs Gaussian filtering and histogram equalization on the pavement images to obtain the target pavement image; a ResNet50 network is used to extract key features of the target pavement image to obtain pavement features; based on the pavement features, a trained defect recognition model is used to perform defect recognition and generate defect recognition results, with an accuracy rate of 95%; and a result output module 14 generates a pavement defect detection report and displays it through a web interface. It should be noted that under clear weather conditions, the pavement defect detection system can complete a road inspection in a specified area within one hour, with a defect recognition accuracy of 95% and a false detection rate of less than 5%.

[0026] In another embodiment (pavement defect detection in adverse weather), a conventional drone system is supplemented with an infrared camera or lidar. The road defect detection steps include: in foggy or nighttime conditions, the drone uses the infrared camera or lidar to assist in collecting road surface images; the collected road surface images are transmitted to a server via a 4G network; a dark channel prior algorithm is applied to remove fog effects to obtain a clear road surface image, or gamma correction is used to enhance the brightness of nighttime images to produce a luminance road surface image; Gaussian filtering and histogram equalization are performed on the clear road surface image or the luminance road surface image to obtain a target road surface image; a trained defect recognition model is used to extract features and identify defects in the target road surface image to generate a defect recognition result; and a defect detection report is generated and displayed based on the defect recognition result. This configuration achieves a 90% defect recognition accuracy in foggy conditions and a 92% defect recognition accuracy in nighttime conditions, significantly outperforming traditional road surface inspection methods.

[0027] In another embodiment (large-scale road network inspection), three drones are deployed simultaneously to cover different road sections. A cloud server centrally processes road image data, implementing SLAM technology to support autonomous drone navigation. SLAM (Simultaneous Localization and Mapping) is a core technology in robotics, autonomous driving, and augmented reality (AR). It uses sensor data to construct real-time environmental maps and estimate its own motion trajectory. The road defect inspection process involves three drones operating simultaneously to cover approximately 100 kilometers of a road network, capturing road images. These images are then transmitted to a cloud server via a 4G / 5G network. The cloud server utilizes distributed computing to parallelize and process the images for improved efficiency. Finally, a comprehensive road defect inspection report is generated, containing defect statistics and maintenance recommendations for all road sections. This road defect detection system can complete a large-scale road inspection within four hours, with a 94% defect identification accuracy rate, providing efficient technical support for large-scale road maintenance.

[0028] Reference Figure 6 , Figure 6 The flow chart of an embodiment of a road surface defect detection method based on deep learning of the present invention is shown. The road surface defect detection method based on deep learning is applied to the road surface defect detection system based on deep learning. The specific implementation steps of the road surface defect detection system based on deep learning of the present invention are further described in detail below. Figure 6 As shown, the method includes steps S110-S140: S110, using an image acquisition module configured in the drone, while the drone flies along a preset path, to acquire an image of the road surface to be inspected, thereby obtaining a road surface image; S120, connecting with the image acquisition module through the image transmission module; S130, connecting to the image transmission module through the image processing module to receive the road surface image transmitted by the image transmission module, and performing preprocessing, feature extraction, and defect recognition on the road surface image in sequence based on a deep learning network to generate a defect recognition result; S140 , connecting the result output module with the image processing module to generate a road surface defect detection report according to the defect recognition result.

[0029] In this embodiment, when the UAV flies along a preset path, it collects images of the road surface to be inspected to obtain a road surface image, and performs preprocessing, feature extraction, and defect recognition on the road surface image in sequence based on a deep learning network to generate a defect recognition result; and generates a road surface defect detection report based on the defect recognition result. It should be noted that the preprocessing, feature extraction, and defect recognition on the road surface image in sequence based on a deep learning network to generate a defect recognition result include: preprocessing the road surface image to obtain a target road surface image; extracting features from the target road surface image through a deep learning network to obtain road surface features; and classifying the road surface features through a trained defect recognition model to obtain the defect recognition result, wherein the defect recognition model is a model obtained by training and verifying a convolutional neural network using data sets under various conditions. It should also be noted that the road surface defect detection report indicates the defect category, the location coordinates of the defect, the severity, defect statistics, and maintenance recommendations; the road surface defect detection report can be displayed through a web interface or a mobile application.

[0030] Furthermore, in this embodiment, the data sets under multiple conditions include sample road surface data sets collected on sunny days, rainy days, foggy days and at night, and the data sets under multiple conditions are used to train and verify the convolutional neural network to obtain the defect recognition model, including: performing data enhancement on the sample road surface data sets collected on sunny days, rainy days, foggy days and at night by data enhancement technology to obtain an enhanced data set, and labeling the enhanced data set, and dividing the labeled enhanced data set into a training data set and a verification data set; inputting the training data set into the convolutional neural network for training until the confidence of the defect category output by the convolutional neural network exceeds a preset confidence; inputting the verification data set into the trained convolutional neural network for verification to output a verification defect category; calculating the recognition defect accuracy rate based on the verification defect category and the defect category in the verification data set; if the recognition defect accuracy rate is greater than the preset accuracy rate, using the trained convolutional neural network as the defect recognition model. It is understandable that if the defect recognition accuracy rate is not greater than the preset accuracy rate, the process returns to executing the step of performing data enhancement on the sample road surface dataset collected on sunny days, rainy days, foggy days, and at night using data enhancement technology to obtain an enhanced dataset, annotating the enhanced dataset, and dividing the annotated enhanced dataset into a training dataset and a validation dataset, until the defect recognition accuracy rate is greater than the preset accuracy rate. It should be noted that in this embodiment, the annotated enhanced dataset is divided into a training dataset and a validation dataset according to a preset ratio, wherein the preset ratio is set according to actual needs, for example, 8:2; the preset accuracy rate is 95%, and in other embodiments, the preset accuracy rate can be 96%, which is set according to actual conditions. It should also be noted that, in this embodiment, both during the model and verification training process and during the road surface defect detection process, the images adopt the standard image data format JPEG format or PNG format, with a standard size of 1920x1080 pixels; the input format of the defect recognition model is the image tensor [batch_size, 3, 224, 224], and the output format includes the defect category and confidence level, such as [crack, 0.95].

[0031] To sum up, in this embodiment, the image acquisition module is used to collect road surface images, and the image processing module is used to preprocess, extract features, and identify defects on the road surface images based on a deep learning network to generate defect identification results; image enhancement technology and auxiliary acquisition devices are used to adapt to harsh environments; the convolutional network is trained with data sets under various conditions to obtain a defect recognition model, thereby improving the robustness of the model, thereby achieving efficient and automated detection of road surface defects. The road surface defect detection system and method based on deep learning in this embodiment are suitable for a variety of practical application scenarios.

[0032] The present invention has been described above in conjunction with the best embodiments, but the present invention is not limited to the embodiments disclosed above, but should cover various modifications and equivalent combinations based on the essence of the present invention.

Claims

1. A road surface defect detection system based on deep learning, characterized in that: include: An image acquisition module is configured in the drone, and is used to acquire an image of the road surface to be inspected when the drone flies along a preset path to obtain a road surface image; An image transmission module connected to the image acquisition module; an image processing module connected to the image transmission module, configured to receive the road surface image transmitted by the image transmission module and sequentially perform preprocessing, feature extraction, and defect recognition on the road surface image based on a deep learning network to generate a defect recognition result; A result output module is connected to the image processing module, and is used to generate a road surface defect detection report based on the defect identification result.

2. The road surface defect detection system based on deep learning according to claim 1, characterized in that: The image processing module includes: An image preprocessing submodule, configured to preprocess the road surface image to obtain a target road surface image; A feature extraction submodule is used to extract features from the target road surface image through a deep learning network to obtain road surface features; The defect recognition submodule is used to obtain the defect recognition result by performing classification based on the road surface characteristics through a trained defect recognition model.

3. The road surface defect detection system based on deep learning according to claim 2, characterized in that: The image preprocessing submodule includes: A filtering submodule, configured to eliminate noise in the road surface image by using Gaussian filtering or median filtering to obtain a filtered road surface image; The clarity adjustment submodule is configured to adjust the clarity of the filtered road image by using a histogram equalization method or an adaptive threshold adjustment technique to obtain the target road image.

4. The road surface defect detection system based on deep learning according to claim 2, characterized in that: The drone is also equipped with an auxiliary acquisition device, and the image processing module further includes: The defogging submodule is used to restore the clarity of the road surface image collected by the auxiliary collection device through a dark road prior algorithm under foggy conditions to obtain a clear road surface image, so that the image preprocessing submodule preprocesses the clear road surface image to obtain the target road surface image.

5. The road surface defect detection system based on deep learning according to claim 4, characterized in that: The image processing module also includes: The brightness adjustment submodule is used to improve the brightness of the road surface image collected by the auxiliary acquisition device under low light conditions through a gamma correction method or a CLAHE algorithm to obtain a brightness road surface image, so that the image preprocessing submodule preprocesses the brightness road surface image to obtain the target road surface image.

6. The deep learning-based road surface defect detection system according to any one of claims 1 to 5, characterized in that: The road surface defect detection system further includes: a first exception handling module, configured to control the drone to return home and save the collected road image when the drone is in an abnormal state; a second exception handling module, configured to cache the road surface image when the transmission of the road surface image is interrupted, and continue transmitting the road surface image after the connection is restored; The third exception handling module is used to issue a prompt for manual review when the defect recognition model makes a classification error.

7. The road surface defect detection system based on deep learning according to claim 1, characterized in that: The road surface defect detection system further includes: The distributed module is used to distribute the road surface images collected by the image acquisition module in the drone in different road sections to different threads when inspecting large-scale road surfaces, so that the threads call the image processing module to process the road surface images in different road sections in parallel, obtain multiple defect identification sub-results, and generate the road surface defect detection report through the result output module based on the multiple defect identification sub-results.

8. A road surface defect detection method based on deep learning, applied to the road surface defect detection system based on deep learning according to any one of claims 1 to 7, characterized in that: The method comprises: The image acquisition module configured in the UAV acquires an image of the road surface to be inspected to obtain a road surface image when the UAV flies along a preset path; Connecting with the image acquisition module via an image transmission module; Connecting to the image transmission module through an image processing module to receive the road surface image transmitted by the image transmission module, and sequentially performing preprocessing, feature extraction, and defect recognition on the road surface image based on a deep learning network to generate a defect recognition result; The result output module is connected to the image processing module to generate a road surface defect detection report based on the defect identification result.

9. The road surface defect detection method based on deep learning according to claim 8, characterized in that: The deep learning network-based preprocessing, feature extraction, and defect recognition of the road surface image are performed sequentially to generate a defect recognition result, including: Preprocessing the road surface image to obtain a target road surface image; Extracting features from the target road surface image using a deep learning network to obtain road surface features; The defect recognition result is obtained by classifying the road surface features using a trained defect recognition model, wherein the defect recognition model is a model obtained by training and verifying a convolutional neural network using data sets under various conditions.

10. The road surface defect detection method based on deep learning according to claim 9, characterized in that: The datasets under various conditions include sample road surface datasets collected on sunny days, rainy days, foggy days, and at night. The defect recognition model is obtained by training and verifying the convolutional neural network using the datasets under various conditions, including: The sample road surface dataset collected on sunny days, rainy days, foggy days and at night is enhanced by data enhancement technology to obtain an enhanced dataset, and the enhanced dataset is labeled, and the labeled enhanced dataset is divided into a training dataset and a validation dataset; Inputting the training data set into the convolutional neural network for training until the confidence level of the defect category output by the convolutional neural network exceeds a preset confidence level; Inputting the verification data set into the trained convolutional neural network for verification to output a verification defect category; Calculating the defect recognition accuracy rate based on the verification defect category and the defect category in the verification data set; If the defect recognition accuracy is greater than the preset accuracy, the trained convolutional neural network is used as the defect recognition model.

Images