Traffic marking damage detection method based on semantic segmentation technology

Abstract

The invention provides a traffic marking damage detection method based on a semantic segmentation technology, and the method comprises the steps: firstly collecting a road image through a camera based on the camera shooting image and semantic segmentation model analysis, carrying out the preprocessing of the collected image, obtaining a damaged marking top view, constructing a marking damage image data set, carrying out the lightweight processing of a DeepLabV3 + network, and carrying out the recognition of the damaged marking. Constructing a marking damage detection semantic segmentation model, and finally training the semantic segmentation model through the image data set to obtain a trained marking damage instance segmentation model; the method provided by the invention has the advantages and positive effects that a semantic segmentation model suitable for accurate detection of multiple types of marked line damage is provided for a damaged marked line image containing multi-semantic feature information, so that accurate identification and contour segmentation of different scale marked line damage types are realized; the problems of low precision of multi-type detection and wrong edge segmentation of a traditional detection mode in the field of marked line damage detection are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision image target detection, and in particular to a traffic marking damage detection method based on semantic segmentation technology. Background Art

[0002] As a key component of traffic facilities, road markings play an important role in transmitting standardized information, guiding and controlling traffic. Relevant departments need to regularly inspect road markings so as to timely maintain and repair severely worn, damaged and aged markings. Traditional road marking damage mainly relies on manual visual inspection, which has low detection efficiency and high false detection rate; laser scanning-based acquisition methods usually require high-performance computers and professional acquisition equipment to be used together, which is difficult to be widely used in road maintenance.

[0003] The computer vision-based method uses cameras to collect data, which is relatively cheap and does not require other supporting equipment. These advantages have made computer vision widely used in road detection. With the continuous emergence of deep neural network models, researchers have combined them with computer vision technology to apply them to road marking recognition and damage detection, which does have obvious advantages in improving the robustness and accuracy of road marking target detection. However, the road marking damage detection method based on deep learning still has obvious shortcomings: First, since the focus of existing technologies is on the accurate positioning of the target marking position, the damaged road marking dataset used for deep learning only marks the position of the damaged road marking, while ignoring the annotation and classification of damaged areas with rich structural information and high feature differences; second, the detection of damaged road marking areas still uses traditional threshold segmentation and edge detection algorithms, relying on simple pixel values ​​or gradient information for segmentation and detection, ignoring the global context information of the image, which is easy to cause false detection or missed detection of damaged edges. Therefore, in view of the defects existing in the existing technology, it is urgent to design a road marking damage detection method that can accurately express the rich structural information and complex geometric rule information after the road marking is damaged, so as to solve the problems of low detection accuracy and edge missegmentation of traditional detection methods in the field of road marking damage detection.

[0004] For damaged road marking images containing multi-semantic feature information, the semantic segmentation algorithm can take into account the global context and high-level feature information of the image, and use a layer-by-layer feature extraction network to extract and optimize the specific target of damaged road marking individuals, which is more suitable for the application needs of multi-scale road marking damage detection research. In addition, image processing technology is used to improve the accuracy of the geometric shape of damaged road markings, and a detailed damaged instance image dataset is established to improve the recognition accuracy and discrimination of small-size samples of semantic segmentation, provide more detailed information for subsequent deep learning training, and ultimately obtain a more accurate damaged segmentation contour effect. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for detecting damaged traffic markings based on semantic segmentation technology, which combines the advantages of deep learning semantic segmentation technology in image processing and the large field of view data information possessed by wide-angle cameras, and realizes the recognition and segmentation of multi-scale and multi-type damaged marking areas.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a traffic marking damage detection method based on semantic segmentation technology, which specifically includes the following steps:

[0007] S1. Acquire a road image through a camera, and pre-process the acquired road image to obtain a top view of a damaged road marking;

[0008] S2, using data annotation software to annotate the damaged instances of the processed damaged road marking image set, and constructing a damaged road marking dataset using the annotated images;

[0009] S3, build a semantic segmentation model with DeepLabV3+ network as the basic model and MobileNetV2 lightweight convolutional neural network as the backbone feature extraction network;

[0010] S4. The damaged features of the damaged marking image are trained and learned through the semantic segmentation model to obtain the final instance segmentation model.

[0011] A further technical solution is that: the method of acquiring road image data through a camera and preprocessing the acquired road image to obtain a top view of a damaged road marking includes:

[0012] Performing distortion calibration on the collected image data through a wide-angle camera to obtain a distortion calibration map after image processing;

[0013] The inverse perspective image transformation is performed on the distortion calibrated image to obtain a bird's-eye view image of the damaged marking area.

[0014] Furthermore, the collected image data is calibrated for distortion by a wide-angle camera, and the distortion calibration map obtained after image processing includes:

[0015] Perform image quality detection on the collected image data and quantitatively evaluate the degree of image distortion;

[0016] Determining a distortion calibration model according to the degree of image distortion;

[0017] Through the following formula:

[0018]

[0019] In the formula, r d Normalized to the radius from the distortion center to the distortion point; r uis the radius of the undistorted point; k i is the radial distortion coefficient;

[0020] The pixel coordinates of the original image are resampled and corrected based on the distortion calibration model to obtain a distortion calibration image after resampling and correction.

[0021] Furthermore, performing inverse perspective image transformation on the image after distortion calibration to obtain a bird's-eye view image of the damaged marking area includes:

[0022] The calibrated image is subjected to denoising, image enhancement and edge segmentation processing to obtain a binary image;

[0023] The cumulative probability Hough transform is used to detect the road segment data, and the optimal marking segment cluster is obtained by filtering and screening according to the segment length and angle;

[0024] The best lane segment cluster is randomly sampled for consistency detection, a fitting model is established, and the estimated position of the lane line vanishing point with the smallest model fitting error (X pos ,Y pos );

[0025] Through the following formula:

[0026]

[0027] Calculate the camera pitch angle θ and yaw angle γ; where (M, N) is the image width and height; α v and α u Respectively represent the vertical and horizontal field of view angle range of the camera;

[0028] According to the camera imaging model and combined with the auxiliary information of camera internal and external parameters required in the inverse perspective transformation process, a bird's-eye view of road markings is obtained.

[0029] A further technical solution is to use data annotation software to annotate the damaged road marking image set after processing, including:

[0030] After slicing, the damaged marking image is cut into sub-images of 710 pixels × 1630 pixels;

[0031] Based on the geometric distribution characteristics of various marking damage locations, the damage types are divided into edge type, internal type, intrusion type and penetration type.

[0032] Use Labelme to semantically segment the sub-images and annotate irregular object instances.

[0033] A further technical solution is to build a semantic segmentation model with DeepLabV3+ network as the basic model and MobileNetV2 lightweight convolutional neural network as the backbone feature extraction network. The model architecture is encoder-decoder, where the encoder includes:

[0034] The backbone feature extraction module MobileNetV2 performs complex feature extraction in high-dimensional space and retains important features in low-dimensional space through expansion and compression operations;

[0035] The atrous spatial pyramid pooling (ASPP) module is formed by the fusion of multi-scale atrous convolution and spatial pyramid pooling. It detects incoming features through multiple convolutional layers with different dilation rates and multiple effective receptive fields, and encodes multi-scale contextual information.

[0036] Furthermore, the multi-scale dilated convolution of the dilated spatial pyramid pooling ASPP module uses a combination of dilated convolutions with dilation rates of 4, 8, 12, and 16 to enhance the model's perception of damage categories of different scales.

[0037] Convolutional layers with dilated convolution rates of 4 and 8 are used to capture smaller details and edge features;

[0038] The convolutional layers with dilated convolution rates of 12 and 16 are used to capture large-area dropouts or larger-scale damage.

[0039] A further technical solution is to train and learn the damage features of the damaged image of the marking line through a semantic segmentation model, including:

[0040] The geometric features of the damaged image are extracted step by step through the backbone feature extraction network. The low-level feature map is directly input into the decoder module, and the deep features are sent to the ASPP module after 16 times downsampling.

[0041] The ASPP module uses four sets of dilated convolution blocks with different expansion rates and a global average pooling block to extract and merge deep features, which are then compressed by a 1×1 convolution block and input into the decoder module.

[0042] The decoder uses a 1×1 convolution kernel to reduce the channel dimension of the low-level feature map and uses a bilinear interpolation method to upsample the high-level feature layer by 4 times;

[0043] The low-level features are concatenated pixel by pixel with the upsampled high-level features, and the features are further fused using a 3×3 convolution block. Finally, linear interpolation upsampling is performed again to gradually restore the original resolution.

[0044] Generate the classification probability of each pixel through the Softmax activation function, select the category with the highest probability as the prediction result, restore the prediction result to the original image size, and output the marking line damage prediction result map;

[0045] During the training process, a multi-classification cross entropy loss function is used to measure the gap between the model prediction results and the actual results, and the model parameters are continuously adjusted based on the calculation results of the total loss function.

[0046] Furthermore, during the model training and evaluation process, the impact of different parameter settings on the detection results is continuously analyzed, including but not limited to the adjustment of the learning rate, the change of the image batch size, and the optimization of the number of iterations.

[0047] Compared with the prior art, the advantages and positive effects of the method provided by the present invention are:

[0048] The deep learning semantic segmentation method is applied to road marking damage detection. The damaged marking annotation data is trained by DeepLabV3+ semantic segmentation model to learn the damaged marking features. The lightweight expansion network MobileNetV2 is adopted as the backbone feature extraction network, and the multi-level expansion coefficient is used in the ASPP module to enhance the model's perception of damage categories of different scales. The decoder converts the instance segmentation task into a pixel-level classification problem and generates an individual distribution map of damaged road markings. For damaged road marking images with multi-semantic feature information, the improved algorithm obtains a more accurate damage segmentation contour effect and realizes the recognition and segmentation of multi-scale and multi-type damaged road marking areas, which is of great significance for the assessment of road marking damage conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of a traffic marking damage detection method based on semantic segmentation technology of the present invention;

[0050] Figure 2 A traffic marking damage type feature map of a traffic marking damage detection method based on semantic segmentation technology of the present invention;

[0051] Figure 3 A schematic diagram of the structure of a DeepLabv3+ semantic segmentation model of a traffic marking damage detection method based on semantic segmentation technology of the present invention;

[0052] Figure 4 It is a schematic diagram of the ASPP network structure of a traffic marking damage detection method based on semantic segmentation technology of the present invention;

[0053] Figure 5 A graph showing the change in loss function during the training process of a traffic marking damage detection method based on semantic segmentation technology of the present invention;

[0054] Figure 6 This is a comparison diagram of various types of damaged original images and segmentation results of a traffic marking damage detection method based on semantic segmentation technology in the present invention. DETAILED DESCRIPTION

[0055] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary implementation modes of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0056] like Figure 1 As shown, according to a traffic marking damage detection method based on semantic segmentation technology according to an embodiment of the present invention, the traffic marking damage detection method based on the improved DeepLabV3+ algorithm includes the following steps:

[0057] S1. Acquire a road image through a camera, and pre-process the acquired road image to obtain a top view of a damaged road marking;

[0058] S2, using data annotation software to annotate the damaged instances of the processed damaged road marking image set, and constructing a damaged road marking dataset using the annotated images;

[0059] S3, build a semantic segmentation model with DeepLabV3+ network as the basic model and MobileNetV2 lightweight convolutional neural network as the backbone feature extraction network;

[0060] S4. The damaged features of the damaged marking image are trained and learned through the semantic segmentation model to obtain the final instance segmentation model.

[0061] Furthermore, in the first step above, the short focal length wide angle camera selected by the camera can capture road marking information in a wider range, but small targets such as road markings will inevitably have image distortion and perspective deformation, so this step also includes the following sub-steps:

[0062] Performing distortion calibration on the collected image data through a wide-angle camera to obtain a distortion calibration map after image processing;

[0063] The inverse perspective image transformation is performed on the distortion calibrated image to obtain a bird's-eye view image of the damaged marking area.

[0064] In some preferred embodiments, the first sub-step specifically includes:

[0065] The checkerboard test card is used to quantitatively evaluate the degree of image distortion, and the inverse transform polynomial distortion model describing the corrected distortion is returned as:

[0066]

[0067] Among them, rd Normalized to the radius from the distortion center to the distortion point; r u is the radius of the undistorted point; k i is the radial distortion coefficient.

[0068] Create a distortion-free, regular virtual pixel array with the image center as the origin and the pixel size as the interval; set the radial distance r of the distortion point d Converted to real radial distance r u , the virtual array is interpolated and resampled at the corresponding correction point to obtain the pixel value. This "reverse mapping" method can avoid the generation of holes or overlaps in the image during the correction process, thereby improving the accuracy of subsequent damaged marking detection.

[0069] The second sub-step specifically includes:

[0070] A world coordinate system is established based on the road plane space to describe the spatial positions of the camera and the object to be measured, and a camera imaging coordinate system is established with the optical center of the lens as the origin to represent the two-dimensional image space;

[0071] Obtain camera internal and external parameters, where the external parameters include: camera pitch angle θ and yaw angle γ, and the height of the camera from the road plane;

[0072] Based on the simplified camera imaging model formula, the two-dimensional image coordinates are converted into three-dimensional world coordinates to obtain a bird's-eye view of the road markings.

[0073] It should be noted that in the road image scene, the camera pitch angle and yaw angle are calculated using the vanishing point coordinates, and the calculation formula is shown in formula (2):

[0074]

[0075] Where (M, N) is the image width and height; α v and α u Respectively represent the vertical and horizontal field of view angle range of the camera;

[0076] Furthermore, the specific steps of obtaining the best estimated vanishing point position based on the continuous iteration of the RANSAC algorithm include:

[0077] The weighted average method and Gaussian filter are used to preprocess the image of the corrected road markings, and the Sobel operator is used to segment the edges of the markings to obtain a binary image.

[0078] The cumulative probability Hough transform is used to randomly select some edge points in the RIO road surface area for voting, and the best line segment cluster N is obtained by filtering according to the line segment length and angle;

[0079] Establish a fitting model, select a fitting function and a distance function; randomly sample from the best line segment data set, and estimate the intersection points based on the randomly sampled line segments;

[0080] Calculate the distance from the estimated vanishing point to other line segments in the line segment data set, and add the line segments with distances less than the threshold to the inner group; estimate the intersection points of the line segments in the inner group again, and calculate the fitting error of the model;

[0081] The inner cluster is used to iterate continuously. After the sampling times are completed, the largest consistent set obtained after sampling is selected to judge the inner and outer points, and finally the vanishing point with the smallest fitting error is selected to estimate the position (X pos ,Y pos ) as the best result.

[0082] In the second step, the present invention proposes the following four types of marking damage based on the geometric distribution characteristics of the damage position according to the collected image set of marking damage examples, such as Figure 2 ,include:

[0083] Edge type: When a vehicle changes lanes or overtakes, the friction between the tire and the road surface causes slight wear on the edge of the marking;

[0084] Internal type: grooves or bulges appear during the marking process, causing the marking damage to spread from the inside to the outside;

[0085] Intrusive: After the edge of the marking is slightly worn, the wear area gradually expands from the outside to the inside under the continuous action of heavy traffic flow;

[0086] Through-type: The adhesion between the marking paint and the road surface is poor, or the internal defects are expanded outward under the action of traffic flow, gradually evolving into large-scale overall shedding, and eventually forming through-type defects.

[0087] After slicing, the image is cut into sub-images of 710 pixels × 1630 pixels, and Labelme is used to perform semantic segmentation and irregular object annotation on the image.

[0088] In the third step, the present invention is based on the DeepLabv3+ network. First, the backbone feature extraction network Xception is replaced by the lightweight convolutional neural network MobileNetV2, and the ASPP module hole convolution rate combination 6, 12, 18 is improved to 4, 8, 12, 16, and a MobileNetV2-DeepLabV3+ semantic segmentation model for multi-scale and multi-type marking damaged area classification detection is proposed. The overall structure of the model is as follows: Figure 3 shown.

[0089] The following table shows the MobileNetV2 network structure configuration, where the stride only represents the stride of the DW convolution in the first bottleneck layer, and the subsequent repeated strides are all 1.

[0090] Table 1 MobileNetV2 network structure configuration:

[0091]

[0092]

[0093] It should be noted that the ASPP module receives feature maps from the backbone network and extracts multi-scale context information in different receptive fields through dilated convolutions with different expansion rates (4, 8, 12, 16), such as Figure 4 As shown;

[0094] Keep a layer of 1×1 convolution to capture local information;

[0095] An additional global pooling layer is added to obtain global features;

[0096] After concatenating the outputs of the above different convolution paths, 1×1 convolution is used to reduce the number of channels.

[0097] Furthermore, the low-level features (higher resolution) extracted by the decoder from the shallow network of MobileNetV2 are fused with the high-level features (lower resolution) output by the ASPP module, including:

[0098] The ASPP output features are fused with shallow features through upsampling bilinear interpolation;

[0099] The high-level features are gradually upsampled by deconvolution or bilinear interpolation to the same resolution as the input image;

[0100] Use a 1×1 convolution to map the decoder output feature map to the number of categories;

[0101] Softmax is used to generate pixel-level probabilities and a segmentation mask is generated based on the maximum probability class.

[0102] During the model training process, the height and width of the input image are defined according to the model requirements, and the target types to be trained are determined to be the four categories shown above;

[0103] Read images and corresponding label information from the input dataset, normalize and resize the images, and select the pre-trained backbone feature extraction network MobileNetV2 to load pre-trained weights;

[0104] Use the cross entropy loss function to measure the gap between the model's predicted results and the actual results, and continuously adjust the model parameters based on the calculation results of the total loss function;

[0105] The more comprehensive and detailed mPA and mIoU are used to evaluate the segmentation accuracy of the model for different disease types of the marking line;

[0106] mPA represents the average value of pixels correctly classified by the model for each category as a percentage of the total pixels, that is:

[0107]

[0108] mIoU is the ratio of the intersection and union of the two sets of true values ​​and predicted values ​​for each category, and the average value is taken after summing, that is:

[0109]

[0110] In the formula, the data set contains k+1 categories, p ij Indicates the number of pixels whose true label value is class i that are incorrectly predicted as class j pixels, p ji On the contrary, p ii Represents the number of correctly predicted pixels of the i-th category.

[0111] Figure 5 The figure shows how the loss function values ​​of the training set and validation set of the model of this method change with the number of training times. Overall, the rate of decrease of the loss function curves of the training set and validation set shows a trend of first fast and then slow, and finally gradually stabilizes; the loss function of the model has a good convergence effect on the road disease marking data set, and the model shows good performance.

[0112] Furthermore, in order to evaluate the effectiveness of the model in the task of road damage segmentation, it was compared with the traditional road damage segmentation method and the deep learning method. After testing, the final results of different evaluation indicators are shown in the following table;

[0113] Table 2 Evaluation indicators of segmentation results of different models:

[0114]

[0115]

[0116] According to the experimental results, compared with traditional methods and deep learning methods, the MobileNetV2-DeepLabV3 model proposed in this paper performs best in four evaluation indicators, including segmentation performance indicators and deployment performance indicators, especially in overall accuracy, which further verifies the effectiveness of realizing refined segmentation of damaged marking images based on the model of the present invention.

[0117] Furthermore, the present invention selected five images of damaged road markings of four damage types with different shapes to evaluate the segmentation performance of the model. The comparison between the original image and the segmentation result is shown in the figure below. Figure 6 As shown in the figure, it can be seen that the model of this method can not only effectively detect the damaged area of ​​the road marking under the road background, but also quickly identify different damage types and accurately segment them.

[0118] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A traffic marking damage detection method based on semantic segmentation technology, characterized in that: The following steps are involved: S1. Acquire a road image through a camera, and pre-process the acquired road image to obtain a top view of a damaged road marking; S2, using data annotation software to annotate the damaged instances of the processed damaged road marking image set, and constructing a damaged road marking dataset using the annotated images; S3. Build a semantic segmentation model with DeepLabV3+ network as the basic model and MobileNetV2 lightweight convolutional neural network as the backbone feature extraction network; S4. The damaged features of the damaged marking image are trained and learned through the semantic segmentation model to obtain the final instance segmentation model.

2. The method for detecting damaged traffic markings based on semantic segmentation technology according to claim 1, characterized in that: The method of acquiring road image data through a camera and preprocessing the acquired road image to obtain an image of a damaged road marking area includes: Performing distortion calibration on the collected image data through a wide-angle camera to obtain a distortion calibration map after image processing; The inverse perspective image transformation is performed on the distortion calibrated image to obtain a bird's-eye view image of the damaged marking area.

3. The method for detecting damaged traffic markings based on semantic segmentation technology according to claim 2, characterized in that: The distortion calibration of the image data collected by the wide-angle camera to obtain the distortion calibration map after image processing includes: Perform image quality detection on the collected image data and quantitatively evaluate the degree of image distortion; Determining a distortion calibration model according to the degree of image distortion; Through the following formula: In the formula, r d Normalized to the radius from the distortion center to the distortion point; r u is the radius of the undistorted point; k i is the radial distortion coefficient; The pixel coordinates of the original image are resampled and corrected based on the distortion calibration model to obtain a distortion calibration image after resampling and correction.

4. The method for detecting damaged traffic markings based on semantic segmentation technology according to claim 2, characterized in that: The step of performing inverse perspective image transformation on the image after distortion calibration to obtain a bird's-eye view image of the damaged marking area includes: The calibrated image is subjected to denoising, image enhancement and edge segmentation processing to obtain a binary image; The cumulative probability Hough transform is used to detect the road segment data, and the optimal marking segment cluster is obtained by filtering and screening according to the segment length and angle; The best lane segment cluster is randomly sampled for consistency detection, a fitting model is established, and the estimated position of the lane line vanishing point with the smallest model fitting error (X pos ,Y pos ); Through the following formula: Calculate the camera pitch angle θ and yaw angle γ; where (M, N) is the image width and height; α v and α u Respectively represent the vertical and horizontal field of view angle range of the camera; According to the camera imaging model and combined with the auxiliary information of camera internal and external parameters required in the inverse perspective transformation process, a bird's-eye view of road markings is obtained.

5. The method for detecting damaged traffic markings based on semantic segmentation technology according to claim 1, characterized in that: The method of using data annotation software to annotate the damaged road marking image set after processing includes: After slicing, the damaged marking image is cut into sub-images of 710 pixels × 1630 pixels; Based on the geometric distribution characteristics of various marking damage locations, the damage types are divided into edge type, internal type, intrusion type and penetration type. Use Labelme to semantically segment the sub-images and annotate irregular object instances.

6. The method for detecting damaged traffic markings based on semantic segmentation technology according to claim 1, characterized in that: The construction is a semantic segmentation model with DeepLabV3+ network as the basic model and MobileNetV2 lightweight convolutional neural network as the backbone feature extraction network. The model architecture is encoder-decoder, where the encoder includes: The backbone feature extraction module is MobileNetV2. The network performs complex feature extraction in high-dimensional space and retains important features in low-dimensional space through expansion and compression operations. The atrous spatial pyramid pooling (ASPP) module is formed by the fusion of multi-scale atrous convolution and spatial pyramid pooling. It detects incoming features through multiple convolutional layers with different dilation rates and multiple effective receptive fields, and encodes multi-scale contextual information.

7. The method for detecting damaged traffic markings based on semantic segmentation technology according to claim 6, characterized in that: The multi-scale dilated convolution of the dilated spatial pyramid pooling ASPP module adopts a combination of dilated convolutions with dilation rates of 4, 8, 12, and 16 to enhance the model's perception of damage categories of different scales. Convolutional layers with dilated convolution rates of 4 and 8 are used to capture smaller details and edge features; The convolutional layers with dilated convolution rates of 12 and 16 are used to capture large-area dropouts or larger-scale damage.

8. The method for detecting damaged traffic markings based on semantic segmentation technology according to claim 1, characterized in that: The training and learning of the damaged features of the damaged marking image by the semantic segmentation model includes: The geometric features of the damaged image are extracted step by step through the backbone feature extraction network. The low-level feature map is directly input into the decoder module, and the deep features are sent to the ASPP module after 16 times downsampling. The ASPP module uses four sets of dilated convolution blocks with different expansion rates and a global average pooling block to extract and merge deep features, which are then input into the decoder module after feature compression through a 1×1 convolution block. The decoder uses a 1×1 convolution kernel to reduce the channel dimension of the low-level feature map and uses a bilinear interpolation method to upsample the high-level feature layer by 4 times; The low-level features are concatenated pixel by pixel with the upsampled high-level features, and the features are further fused using a 3×3 convolution block. Finally, linear interpolation upsampling is performed again to gradually restore the original resolution. Generate the classification probability of each pixel through the Softmax activation function, select the category with the highest probability as the prediction result, restore the prediction result to the original image size, and output the marking line damage prediction result map; During the training process, a multi-classification cross entropy loss function is used to measure the gap between the model prediction results and the actual results, and the model parameters are continuously adjusted based on the calculation results of the total loss function.

9. The method for detecting damaged traffic markings based on semantic segmentation technology according to claim 8, characterized in that: During the model training and evaluation process, the impact of different parameter settings on the detection results is continuously analyzed, including but not limited to the adjustment of the learning rate, the change of the image batch size, and the optimization of the number of iterations.

Images