CN114066808B - Pavement defect detection method and system based on deep learning - Google Patents

Abstract

The invention relates to a pavement defect detection method and system based on deep learning, wherein the method comprises the following steps: constructing an image segmentation model, wherein a U-MDN uses a U-Net feature extraction network as a main body, extracts multi-scale features obtained from different convolution stages, uses a U-MDM to further extract deep features of defects, and then efficiently fuses the multi-scale features obtained from shallow features and deep features to obtain a prediction result; and (5) quantitatively evaluating the pavement defects, and further calculating pavement damage indexes based on quantitative parameters to obtain corresponding pavement damage grades. The method overcomes the defect of single characteristic of common standard convolution extraction, and can remove redundant information and reduce the calculated amount; the improved deep neural network is used for detecting the pavement defects in real time, so that the defects of the traditional manual pavement detection method are overcome, and the detection efficiency is improved.

Description

A road defect detection method and system based on deep learning

Technical Field

The present invention relates to the technical field of road surface defect detection, and in particular to a road surface defect detection method and system based on deep learning.

Background Art

With the rapid development of the transportation industry, my country has formed a road network of a certain scale. The large number of surface defects generated during use indicates that the existing roads have entered the stage of large-scale inspection and maintenance. The increasing diversity and complexity of the road network has led to higher and higher performance requirements for defect detection methods in this field. Therefore, being able to detect cracks in a timely manner and more accurately and quickly is of great significance to road maintenance and driving safety.

Pavement cracks are one of the most common diseases, posing a huge threat to highway safety. Being able to detect, repair and prevent further deterioration in a timely manner has become an important task for the transportation department. Early pavement defect detection methods mainly include video measurement, ground penetrating radar, laser measurement and infrared measurement. These methods usually use manual evaluation, obtain data on the spot and then return to the laboratory for data analysis. This evaluation method not only increases costs, is time-consuming and labor-intensive, but also has low detection accuracy and does not meet the requirements. Later, research on pavement defect detection technology focused on machine learning, including detection methods such as oriented gradient histogram, support vector machine, Canny edge detection and Otsu threshold segmentation. This type of threshold-based detection method has a good detection effect under simple backgrounds. However, nowadays, the complexity of road conditions continues to deepen, and some factors such as oil stains, shadows and foreign matter seriously affect the detection accuracy and restrict the development of image segmentation technology. Therefore, it cannot meet the actual detection requirements.

With the development of deep learning, various deep neural networks have been used in the field of road defect detection. They can mine deeper features and filter out the influence of complex background to a large extent, becoming the mainstream trend in the field of road defect detection. The typical fully convolutional network (FCN) has become a milestone segmentation network in the field of semantic segmentation due to its good segmentation effect achieved by its skip-level fusion structure. However, FCN treats different features such as crack edges, patterns and shapes obtained in the downsampling stage in the same way, which has a great impact on the segmentation results. Moreover, for relatively thin targets such as road cracks, the continuous convolution operation of FCN will cause serious loss of local details. In order to improve this problem, many researchers at home and abroad have improved FCN and created networks such as U-Net and SegNet. These networks follow the basic idea of the FCN codec structure. The difference is that both of them fuse the feature map obtained in the upsampling stage with the underlying features obtained in the downsampling stage, which increases the accuracy of positioning and the integrity of semantic information, and improves the crack segmentation effect. However, these networks based on codec structures are only ideal for some cracks with good continuity and large contrast with strong interference. Therefore, the detection effect is still poor when the light changes greatly, the background interference is strong, and the boundaries between some fine cracks and noise are blurred. It can be seen that achieving high-precision crack detection is still a research focus.

In summary, it is very necessary to seek a road defect detection method to solve the above problems.

Summary of the invention

The purpose of the present invention is to provide a road defect detection method and system based on deep learning. By setting U-MDM, the module adopts a 4-scale up and down sampling structure and combines the advantages of dilated convolution, overcoming the disadvantage of single feature extraction of commonly used standard convolution, and can remove redundant information to reduce the amount of calculation.

In order to solve the above technical problems, the technical solutions of the road surface defect detection method and system based on deep learning provided by the present invention are as follows:

In a first aspect, an embodiment of the present invention discloses a road surface defect detection method based on deep learning, the method comprising the following steps:

Image segmentation model construction, which includes a U-shaped multi-scale dilated network, which embeds a U-shaped multi-scale dilated convolution module into a U-Net deep neural network before upsampling to form a U-MDN. The U-MDN uses the U-Net feature extraction network as the backbone to extract multi-scale features obtained from different convolution stages, and uses the U-MDM to further extract deep features of defects, and then efficiently fuses the multi-scale features obtained from shallow features and deep features to obtain prediction results;

The pavement defect quantitative assessment is to quantify the defect parameters according to the obtained prediction results, and further calculate the pavement damage index based on the quantified parameters to obtain the corresponding pavement damage grade.

In any of the above solutions, preferably, the U-Net deep neural network includes a symmetrical U-shaped structure including a down-sampling image compression path and an up-sampling image expansion path.

In any of the above schemes, preferably, four Convolution block1s and one Convolution block2 are used in the downsampling stage, four Convolution block3s are used in the upsampling stage, each convolution block in Convolution block1 and 3 uses two convolution operations with a convolution kernel of 3*3 and one MaxPooling operation of 2*2, and three convolution operations with a convolution kernel of 3*3 are used in Convolution block2.

In any of the above schemes, preferably, the U-MDN network consists of 4 Convolution block1s, 1 Convolution block2, 1 U-MDM and 4 Convolution block3s, the U-MDN network input image is 256*256*3, and the output feature is a black and white binary image of 256*256*2.

In any of the above schemes, preferably, the image segmentation model construction includes model training. During the training process, the hyperparameter Batch size is set to 8, 150 epochs are set for each training, the initial learning rate lr is 1e-4, the Adam optimizer is used to converge the network, and Adam's momentum and adaptive learning rate are used to speed up the convergence speed.

In any of the above solutions, preferably, the quantitative calculation of the quantization parameter of the defect comprises the following steps:

Calculate the conversion ratio based on the number of crack image pixels N and the actual size D The actual crack size of the segmented crack image is D=p·N. The crack contour is extracted from the segmented binary image, and the crack area is filled with an inscribed circle.

In any of the above schemes, preferably, the quantitative calculation of the quantization parameter of the defect includes the following steps: h is the length of the outer contour of the crack area, w is the width of the outer contour of the crack area, the centers of adjacent circles are connected, and the distance between the centers is calculated, w _i is the difference between the horizontal coordinates of the center pixels of two adjacent circles, _hi is the difference between the vertical coordinates of the center pixels of two adjacent circles, then l _i is the approximate actual distance of the crack, then the distance between the n center points is the actual length L of the entire crack, the actual width R is the average value of the diameter of the inscribed circle, and the calculation method of the quantization parameter width to be calculated is: The length is calculated as: The area of the circumscribed rectangle is calculated as: S = p ² ·h·w.

Preferably, in any of the above schemes, the U-MDN performs MaxPooling operations with Pooling sizes of 16*16, 8*8, 4*4 and 2*2 on the obtained 16*16*512 feature map to obtain feature maps of 1*1*512, 2*2*512, 4*4*512 and 8*8*512 respectively, and then performs convolution operations with expansion rates of 1, 2, 4 and 6 to obtain feature maps of 1*1*256, 2*2*256, 4*4*256 and 8*8*256.

In any of the above schemes, it is preferred that the feature map is restored to 256 feature maps with a height of 16 pixels and a width of 16 pixels after deconvolution, and the deconvolution kernel sizes are 16*16, 8*8, 4*4 and 2*2 respectively. Four 16*16*256 are superimposed and fused to obtain a feature map of 16*16*1024, and then a convolution operation is performed to finally obtain the output feature map of U-MDM with a size of 16*16*512. After U-MDM, four Convolution block3s are performed to obtain 32 feature maps with a height of 256 pixels and a width of 256 pixels. After the Sigmoid activation function, a 256*256*2 road defect detection image is obtained, and each image is a black and white binary image, which is divided into two categories: defect area and background area, where the white area represents the defect and the black area represents the background.

Compared with the prior art, the present invention has the following beneficial effects:

By adding U-MDM, the module adopts a 4-scale up and down sampling structure and combines the advantages of dilated convolution, overcoming the disadvantage of single feature extraction of commonly used standard convolution, removing redundant information and reducing the amount of calculation; by using an improved deep neural network to perform real-time detection of pavement defects, it overcomes the shortcomings of traditional manual pavement detection methods and improves detection efficiency; the pavement condition health assessment is performed on the pavement defect detection results, providing sufficient data support for subsequent highway maintenance management.

In a second aspect, a road surface defect detection system based on deep learning includes:

A construction module for image segmentation model construction, wherein the image segmentation model construction includes a U-shaped multi-scale dilated network, wherein the U-shaped multi-scale dilated convolution module is embedded into the U-Net deep neural network before upsampling to form a U-MDN, wherein the U-MDN uses the U-Net feature extraction network as a backbone to extract multi-scale features obtained from different convolution stages, and uses the U-MDM to further extract deep features of defects, and then efficiently fuses the multi-scale features obtained from shallow features and deep features to obtain prediction results;

The evaluation module is used for quantitative evaluation of road surface defects. The quantitative evaluation of road surface defects quantitatively calculates the quantitative parameters of the defects according to the obtained prediction results, and further calculates the road surface damage index based on the quantitative parameters to obtain the corresponding road surface damage grade.

Compared with the prior art, the present invention has the following beneficial effects: by adding U-MDM, the module adopts a 4-scale up-down sampling structure and combines the advantages of dilated convolution, overcoming the disadvantage of single feature extraction of commonly used standard convolution. It can also remove redundant information and reduce the amount of calculation; by using an improved deep neural network to detect road defects in real time, it overcomes the shortcomings of traditional manual road detection methods and improves detection efficiency; the road condition health assessment is performed on the road defect detection results, providing sufficient data support for subsequent highway maintenance management.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings are used for further understanding of the present invention and are used to explain the present invention together with the embodiments of the present invention, but do not constitute a limitation of the present invention.

FIG1 is a block diagram of the overall structure of a road surface defect detection method based on deep learning according to the present invention;

FIG2 is a schematic diagram of a convolution block of a road surface defect detection method based on deep learning according to the present invention;

FIG3 is a schematic diagram of the structure of a road surface defect detection method U-MDN based on deep learning according to the present invention;

FIG4 is a schematic diagram of a U-MDN network structure of a road surface defect detection method based on deep learning according to the present invention.

FIG5 is a schematic diagram of calculating quantitative parameters of a road surface defect detection method based on deep learning according to the present invention.

FIG6 is a schematic diagram of a road surface defect detection system based on deep learning according to the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

A road surface defect detection method based on deep learning, specifically including: image segmentation model construction and road surface defect quantitative assessment. The present invention regards road surface defect detection as a binary classification task, and extracts features of road surface defects by improving U-Net. Each pixel in the defect image is divided into two categories: defect and background, so as to obtain accurate defect quantification information. Specifically, the present invention designs a U-type multi-scale dilated convolution module (U-Multiscale Dilated Module, U-MDM). U-MDM is embedded into U-Net before upsampling to form a U-type multi-scale dilated network (U-Multiscale Dilated Network, U-MDN). The network uses the U-Net feature extraction network as the backbone to extract multi-scale features obtained from different convolution stages. Then, U-MDM is used to further extract deep features of defects, and then the multi-scale features obtained from shallow features and deep features are efficiently fused to finally obtain prediction results.

There are 5 convolution blocks in the downsampling stage of the feature extraction network. The defect images of different resolutions are unified into 3-channel, 256-pixel-high, 256-pixel-wide RGB images and then input into the network. After downsampling, an image with 512 channels, 16 pixels in height, and 16 pixels in width is obtained. The 16*16*512 feature layer is then input into U-MDM, and after 4-scale maximum pooling and dilated convolution, it is fused to obtain a 16*16*512 feature layer again. Finally, the 4 convolution blocks upsampled by U-Net obtain a 256*256*32 feature layer, which is then passed through the Sigmoid function to obtain a road defect prediction image of size 256*256*2 containing only defects and background.

After the segmentation model is built, the model is trained using the data set and hyperparameters are set. After the training is completed, the trained model is used to detect defect images. Finally, a defect mask image is generated in real time for subsequent defect quantitative assessment.

After the pavement defect image is segmented, the extracted defect targets need to be quantified for efficient subsequent highway maintenance management. Various defects are quantified based on the prediction results, including the length, width and direction of various linear cracks, the coverage area of reticular cracks and the density of cracks, the geometric area of defects such as potholes, protrusions and depressions, and quantitative parameters such as the length and width of the circumscribed rectangle. The pavement damage index is further calculated based on these quantitative parameters, and the corresponding pavement damage grade is finally obtained.

According to the relevant highway condition assessment standards, when evaluating the pavement condition, two indicators, namely damage type and damage degree, are generally used to assess it. Pavement damage is mainly divided into several types, such as horizontal and vertical linear cracks, irregular cracks such as mesh and block cracks, rutting, depressions and convexities, and potholes. The degree of damage can be divided into three levels according to the width of the crack: mild, moderate, and severe. It can be seen that the severity of damage is also different for different crack types. In order to unify the standards, the pavement condition index PCI is used to evaluate it. The range of PCI is [0,100], which reflects the severity of pavement damage. The smaller the value, the more serious the pavement damage.

In order to better understand the above technical solution, the technical solution of the present invention will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

Example:

The U-MDN designed by the present invention mainly consists of two parts in the overall network structure: U-Net and U-MDM. U-Net is first used to extract multi-scale information obtained from different convolution blocks, and then U-MDM is used to further extract deep features from four different scales. Finally, U-Net upsampling is used to merge shallow features with deep features. Finally, the prediction result is obtained through the Sigmoid function.

Segmentation network design, U-Net deep neural network is a symmetrical U-shaped structure that includes a down-sampled image compression path and an up-sampled image expansion path. This structure follows the basic idea of the FCN codec structure. And both networks use convolution operations instead of full connection operations, which can greatly reduce the number of network model parameters and improve the model training speed. The difference is that U-Net focuses on the upsampling stage, increases the number of feature maps in the upsampling stage, and then fuses it with the underlying features obtained by downsampling, which increases the accuracy of positioning and the integrity of semantic information, making the network more generalized. Since the downsampling process in this structure can capture rich semantic information and the upsampling process can perform precise positioning, good results can be obtained using a small amount of data. Therefore, the present invention uses U-Net as the backbone of the feature extraction network to extract semantic information obtained from different convolution blocks in the downsampling process.

The convolution blocks used in U-Net are shown in Figure 2. (a) and (b) are convolution blocks in the downsampling stage, and (c) is a convolution block in the upsampling stage. Four convolution blocks 1 and one convolution block 2 are used in the downsampling stage, and four convolution blocks 3 are used in the upsampling stage. Each convolution block in Convolution block 1 and 3 uses two convolution operations with a convolution kernel of 3*3 and one MaxPooling operation with a convolution kernel of 2*2, while three convolution operations with a convolution kernel of 3*3 are used in Convolution block 2. Batch Normalization (BN) is added between the convolution layer and the activation function of each convolution block. Using BN will increase the gradient and avoid the problem of gradient disappearance. The increase in gradient means that the learning convergence speed is accelerated, which greatly speeds up the model training speed. The activation function uses ReLU and Sigmoid functions. The Sigmoid activation function is used in the last layer of the network, while the ReLu activation function is used in other convolution layers. The ReLu activation function has better effects than the Sigmoid function and converges faster. It is suitable for solving the problem of vanishing gradients in deep networks. The Sigmoid function can limit the calculation results to the range of [0,1], making it easier to use probability to distinguish between targets and backgrounds. It is more suitable for binary classification tasks of images and is often used in the last layer of the network. The calculation formulas for the ReLu and Sigmoid functions are shown in (1) and (2) respectively.

Although neural networks based on codec structures can extract multi-scale information of cracks, this multi-scale information is not enough to detect cracks from some complex backgrounds. Therefore, inspired by the multi-scale dilation module (MDM) used by other researchers, and considering that U-Net only uses one standard convolution filter to extract crack features, only one contextual information of the crack can be obtained. For complex crack images with different topological structures such as width, length and direction, using only one convolution layer is not conducive to the extraction of multiple effective features. Moreover, continuous convolution operations have a good effect on the recognition of targets with high recognition, but are not friendly to some targets with low recognition. Dilated convolution can increase the receptive field without losing information, thereby effectively extracting richer semantic information and enhancing the recognition ability of target areas with weak resolution. Therefore, the U-type multi-scale dilated convolution module (U-MDM) is designed in combination with the characteristics of dilated convolution.

The U-MDM network structure is shown in Figure 3. This module integrates deep feature information of four different scales. Specifically, U-MDM does not directly fuse the feature layer obtained after U-Net downsampling, but first undergoes a maximum pooling operation with four different scales of Pooling size 16, 8, 4, and 2, which can remove redundant information and reduce the amount of calculation. It then passes through a convolution block. Unlike the convolution block in U-Net, the expansion rates of the convolution operation in this convolution block are 1, 2, 4, and 6, and the convolution kernels are all 3*3, and BN and ReLu are added after the convolution. Finally, after upsampling and restoring the size, it is input into the U-Net upsampling structure for fusion. This forms a U-shaped structure that helps to further extract deeper features of the crack.

The overall structure diagram of U-MDN is shown in Figure 4. The network consists of 4 Convolution blocks 1, 1 Convolution block 2, 1 U-MDM and 4 Convolution blocks 3. The network input image is 256*256*3 (256 pixels wide, 256 pixels high, 3 channels, and width*height*number of channels in the following), and the output feature is a 256*256*2 black and white binary image. Specifically, the input 256*256*3 image passes through 4 Convolution blocks 1 in succession to obtain 256 feature maps with a height of 16 pixels and a width of 16 pixels. After passing through 1 Convolution block 2, 512 feature maps with a height of 16 pixels and a width of 16 pixels are obtained. After passing through U-MDM, 512 feature maps with a height of 16 pixels and a width of 16 pixels are still obtained. The difference is that this feature map integrates four deep features of different scales. First, U-MDM performs MaxPooling operations with Pooling sizes of 16*16*512, 8*8, 4*4 and 2*2 on the 16*16*512 feature map to obtain 1*1*512, 2*2*512, 4*4*512 and 8*8*512 feature maps respectively. Then, it undergoes convolution operations with expansion rates of 1, 2, 4 and 6 to obtain feature maps of 1*1*256, 2*2*256, 4*4*256 and 8*8*256. After that, it undergoes deconvolution to restore the size to 256 feature maps with a height of 16 pixels and a width of 16 pixels. The deconvolution kernel sizes are 16*16, 8*8, 4*4 and 2*2 respectively. Finally, the four 16*16*256 images are superimposed and fused to obtain a 16*16*1024 feature map, and then a convolution operation is performed to finally obtain the output feature map of U-MDM, with a size of 16*16*512. After U-MDM, four Convolution blocks3 are performed to obtain 32 feature maps with a height of 256 pixels and a width of 256 pixels. Finally, a 256*256*2 road defect detection image is obtained through the Sigmoid activation function. Each image is a black and white binary image, which is divided into two categories: defect area and background area. The white area represents the defect and the black area represents the background.

During model training, the hyperparameter Batch size is set to 8, 150 epochs are set for each training, the initial learning rate lr is 1e-4, and the Adam optimizer is used to converge the network. Adam's momentum and adaptive learning rate are used to speed up the convergence. Adam replaces the classic stochastic gradient descent (SGD) method, which can update the network weights more effectively.

In deep neural networks, cross entropy is often used as a loss function for handling classification problems. Therefore, the training process uses the cross entropy loss function, which is widely used in neural network models, to calculate the loss value (Loss) to describe the error between the model's prediction results and the true label during the training process. The loss function used in the study is a type of cross entropy, Binarycrossentropy (see Formula 3). This loss function is often used to handle binary classification problems, and it involves calculating class probabilities during use, so it is usually used together with the sigmoid function in Formula 1.

Where N is the output size, n∈[1,N], y _n ∈{0,1} is the true label, is the predicted probability.

Quantitative assessment of pavement defects. In the task of pavement defect detection, the geometric parameters of different defects are calculated to provide data support for subsequent pavement repair work. Pavement defect quantification calculates parameters such as the length and width of unidirectional cracks and the size and area of reticular cracks. The characteristic parameters that need to be calculated for cracks are width, length, and circumscribed rectangular area.

First, the conversion ratio is calculated based on the number of crack image pixels N and the actual size D Then the actual crack size of the segmented crack image is D=p·N. Secondly, extract the crack contour of the segmented binary image and fill the crack area with an inscribed circle. Please refer to Figure 5, h is the length of the outer contour of the crack area, and w is the width of the outer contour of the crack area. Connect the centers of adjacent circles and calculate the distance between the centers. w _i is the difference in the horizontal coordinates of the center pixels of two adjacent circles, and h _i is the difference in the vertical coordinates of the center pixels of two adjacent circles. Then l _i is the approximate actual distance of the crack. Then the distance between the n center points is the actual length L of the entire crack, and the actual width R is the average value of the inscribed circle diameter. The quantitative parameters that need to be calculated and their practical significance are as follows:

(1) Width

Its physical meaning is the distance between the vertical line at a point on the crack boundary and the intersection of the crack, which reflects the degree of crack opening.

(2) Length

Its physical meaning is the length of the crack skeleton after refinement, which reflects the extent of crack extension.

(3) Area of the circumscribed rectangle

S＝p ² ·h·w (7)

Its physical meaning is the circumscribed rectangular area of cracks of various shapes, which reflects the degree of crack coverage.

Among the road surface defects, two indicators, damage type and damage degree, are generally used to evaluate them. After obtaining quantitative information based on the calculation formulas of various quantitative parameters, the damage degree of various defects can be obtained according to the classification standards. According to the relevant highway condition evaluation indicators, common road surface defects and their evaluation standards are listed. See Table 2.

Table 1 Pavement defect types and damage classification

It can be seen that the severity of damage of different crack types is also different. In order to unify the standard, the pavement condition index (PCI) is used to evaluate it. The PCI value reflects the severity of pavement damage. The smaller the value, the more serious the pavement damage. The various indicators and maintenance suggestions are shown in Table 3.

Table 2 Evaluation indexes of pavement damage and maintenance suggestions

The value range of PCI is [0,100] and is calculated using formulas 4 and 5.

Among them, DR is the pavement damage rate (%); _Ai is the cumulative area of the i-th type of pavement damage (m ² ); A is the total area of the investigated road section (m ² ); _ωi is the damage weight value of the i-th type of pavement damage; _a0 is 15.00 for asphalt pavement and 10.66 for concrete pavement; _ai is 0.412 for asphalt pavement and 0.461 for concrete pavement; _i0 is the total number of damage types including the damage degree. Finally, the pavement damage condition grade is obtained according to the PCI value.

In a second aspect, a road surface defect detection system based on deep learning includes:

A construction module for image segmentation model construction, wherein the image segmentation model construction includes a U-shaped multi-scale dilated network, wherein the U-shaped multi-scale dilated convolution module is embedded into the U-Net deep neural network before upsampling to form a U-MDN, wherein the U-MDN uses the U-Net feature extraction network as a backbone to extract multi-scale features obtained from different convolution stages, and uses the U-MDM to further extract deep features of defects, and then efficiently fuses the multi-scale features obtained from shallow features and deep features to obtain prediction results;

The evaluation module is used for quantitative evaluation of road surface defects. The quantitative evaluation of road surface defects quantitatively calculates the quantitative parameters of the defects according to the obtained prediction results, and further calculates the road surface damage index based on the quantitative parameters to obtain the corresponding road surface damage grade.

The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims (2)

1. A road surface defect detection method based on deep learning, characterized in that: the method comprises the following steps: Image segmentation model construction, which includes a U-shaped multi-scale dilated network, which embeds a U-shaped multi-scale dilated convolution module into a U-Net deep neural network before upsampling to form a U-MDN. The U-MDN uses the U-Net feature extraction network as the backbone to extract multi-scale features obtained from different convolution stages, and uses the U-MDM to further extract deep features of defects, and then efficiently fuses the multi-scale features obtained from shallow features and deep features to obtain prediction results; Quantitative assessment of road surface defects, the quantitative assessment of road surface defects quantifies and calculates the quantitative parameters of defects according to the obtained prediction results, and further calculates the road surface damage index based on the quantitative parameters to obtain the corresponding road surface damage level; the U-Net deep neural network includes a symmetrical U-shaped structure of a down-sampling image compression path and an up-sampling image expansion path; 4 Convolution block1s and 1 Convolution block2 are used in the down-sampling stage, and 4 Convolution block3s are used in the up-sampling stage. Each convolution block in Convolution block1 and 3 uses two convolution operations with a convolution kernel of 3*3 and one MaxPooling operation of 2*2, and three convolution operations with a convolution kernel of 3*3 are used in Convolution block2; the U-MDN network consists of 4 Convolution block1s, 1 Convolution block2, 1 U-MDM and 4 Convolution block3s. The U-MDN network input image is 256*256*3, and the output feature is a black and white binary image of 256*256*2; the image segmentation model construction includes model training. During the training process, the hyperparameter Batch is set The size is 8, 150 epochs are set for each training, the initial learning rate lr is 1e-4, the Adam optimizer is used to converge the network, and the momentum and adaptive learning rate of Adam are used to speed up the convergence speed; the quantitative calculation of the quantization parameters of the defects includes the following steps: Calculate the conversion ratio based on the number of crack image pixels N and the actual size D The actual crack size of the crack image after segmentation is D=p·N. The crack contour is extracted from the binary image after segmentation, and the crack area is filled with an inscribed circle. The quantitative calculation of the quantization parameters of the defect includes the following steps: h is the length of the outer contour of the crack area, w is the width of the outer contour of the crack area, the centers of adjacent circles are connected, and the distance between the centers is calculated. w _i is the difference between the horizontal coordinates of the center pixels of two adjacent circles, _hi is the difference between the vertical coordinates of the center pixels of two adjacent circles, then l _i is the approximate actual distance of the crack, then the distance between the n center points is the actual length L of the entire crack, the actual width R is the average value of the diameter of the inscribed circle, and the calculation method of the quantization parameter width to be calculated is: The length is calculated as: The calculation method of the circumscribed rectangle area is: S = p ² ·h·w; the U-MDN performs MaxPooling operations with Pooling sizes of 16*16*512, 8*8, 4*4 and 2*2 on the obtained 16*16*512 feature map to obtain 1*1*512, 2*2*512, 4*4*512 and 8*8*512 feature maps respectively, and then performs convolution operations with expansion rates of 1, 2, 4 and 6 to obtain 1*1*256, 2*2*256, 4*4*256 and 8*8*2 56 feature maps; the feature map is restored to 256 feature maps with a height of 16 pixels and a width of 16 pixels after deconvolution, and the deconvolution kernel sizes are 16*16, 8*8, 4*4 and 2*2 respectively. Four 16*16*256 feature maps are superimposed and fused to obtain a 16*16*1024 feature map, and then a convolution operation is performed to finally obtain the output feature map of U-MDM with a size of 16*16*512. After U-MDM, four Convolution block3s are performed to obtain 32 feature maps with a height of 256 pixels and a width of 256 pixels. After the Sigmoid activation function, a 256*256*2 road defect detection image is obtained. Each image is a black and white binary image, which is divided into two categories: defect area and background area, where the white area represents the defect and the black area represents the background. 2. A road surface defect detection system based on deep learning, characterized in that: using the method according to claim 1, comprising: A construction module for image segmentation model construction, wherein the image segmentation model construction includes a U-shaped multi-scale dilated network, wherein the U-shaped multi-scale dilated convolution module is embedded into the U-Net deep neural network before upsampling to form a U-MDN, wherein the U-MDN uses the U-Net feature extraction network as a backbone to extract multi-scale features obtained from different convolution stages, and uses the U-MDM to further extract deep features of defects, and then efficiently fuses the multi-scale features obtained from shallow features and deep features to obtain prediction results; The evaluation module is used for quantitative evaluation of road surface defects. The quantitative evaluation of road surface defects quantitatively calculates the quantitative parameters of the defects according to the obtained prediction results, and further calculates the road surface damage index based on the quantitative parameters to obtain the corresponding road surface damage grade.