CN-122023417-A - Road surface disease detection method, system and equipment

CN122023417ACN 122023417 ACN122023417 ACN 122023417ACN-122023417-A

Abstract

The invention discloses a road pavement disease detection method, a system and equipment, which relate to the technical field of road pavement image processing and have the technical scheme that road pavement images to be detected are obtained, the road pavement images to be detected are input into a pre-trained disease area segmentation model to be segmented and extracted to obtain disease area images of road pavement, wherein a first generation reactance network is formed according to a pre-trained first discriminator and a first generator, the first generation reactance network is trained, a disease detection model is obtained when the training times reach the maximum times, the disease area images are input into the pre-trained disease detection model to be detected to obtain disease detection results of the road pavement, the pre-built YOLOv neural network is trained, and the disease detection model is obtained when a first loss function converges. The problems that detailed information cannot be captured and diseases are missed when road surface diseases are detected in the prior art are solved.

Inventors

XIN SHUJIE
Yi Zaishu
YANG ZHI
WEI SHUGANG
LI XU
Han xianke
ZHOU RONGZHENG
KANG YI
CHU ZHIJIAN
Zang Chaoyin
Lei Chunxiu
ZHAO QIN

Assignees

四川省交通运输发展战略和规划科学研究院
甘孜藏族自治州交通运行监测中心

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (10)

1. A method for detecting road surface diseases, which is characterized by comprising the following steps: Acquiring a road surface image to be detected of a road; Inputting a pavement image to be detected into a pre-trained disease area segmentation model for segmentation extraction to obtain a disease area image of a pavement, wherein Unet networks comprising encoders and decoders are used as repeated units to stack to obtain a first generator, a first generation reactance network is formed according to the pre-trained first discriminator and the first generator, the first generation reactance network is trained, and when the training times reach the maximum times, the first generator with segmentation extraction capacity for pavement diseases of the pavement image is obtained and is used as the disease area segmentation model; And inputting the disease area image into a pre-trained disease detection model to detect, thereby obtaining a disease detection result of the highway pavement, wherein the pre-built YOLOv neural network is trained, and the disease detection model is obtained when the first loss function converges.
2. The method according to claim 1, wherein the output part of the encoder is configured with a high-frequency feature extraction branch and a low-frequency feature extraction branch in parallel, wherein the high-frequency feature extraction branch is used for extracting high-frequency information of different-scale feature maps output by the encoder, and the low-frequency feature extraction branch is used for extracting low-frequency information of different-scale feature maps output by the encoder; And the output parts of the high-frequency characteristic extraction branch and the low-frequency characteristic extraction branch are provided with a characteristic fusion module, and the output of the characteristic fusion module is used as the input of the decoder.
3. The method for detecting road surface diseases according to claim 2, wherein the training of the first generation-oriented network is as follows: constructing a second generation countermeasure network by a second generator and a pre-trained second discriminator, and training the second generator according to the first generation countermeasure network to obtain a trained second generator, wherein the pre-training of the second discriminator is completed by disease images of the same disease type under different distribution characteristics; The method comprises the steps of forming a first generation countercheck network by a first generator and a first discriminant which is trained in advance, loading training parameters of a second generator which is trained in all decoders of repeated UNet network units of the first generator, training the first generator after the training parameters are loaded according to the first countercheck network, and obtaining the first generator with segmentation and extraction capacity on diseases of a disease area image when the training parameters reach the maximum number of times, wherein the training of the first discriminant is completed by data marked by the disease types of real pavement.
4. A method for detecting road surface diseases according to claim 3, wherein a composite function of fusion of Dice and weighted two-class cross entropy is used as the loss function of the first generator training.
5. The method for detecting road surface diseases according to claim 1, wherein YOLOv neural network means adding GAM module to output of C2f module of backbone network part in initial YOLOv network, and replacing backbone network part space pyramid pooling module in initial YOLOv network with feature interaction module; The first loss function is a weighted sum function of the bounding box regression loss and the weighted cross entropy loss.
6. The method of claim 5, wherein the GAM module comprises a channel attention module and a spatial attention module; The channel attention module is used for carrying out channel attention calculation on a first feature map of the road surface image extracted from the main network to obtain a channel attention feature map, and multiplying the channel attention feature map with the first feature map to obtain a second feature map; and performing spatial attention calculation on the second feature map through the spatial attention module to obtain a spatial attention feature map, and multiplying the spatial attention feature map by the second feature map to obtain a third feature map.
7. The method of claim 1, wherein the feature interaction module is formed by stacking three different-scale transducer encoders, and wherein each different-scale transducer encoder input is the output of the GAM module.
8. The method for detecting road surface diseases according to claim 7, wherein the transducer encoder comprises a two-dimensional sine and cosine position coding layer and a linear weight layer, and the two-dimensional sine and cosine position coding layer and the linear weight layer are fused through a multiplication operator.
9. A highway pavement defect detection system, comprising: The image acquisition unit is used for acquiring road images to be detected of the road; The pavement disease segmentation unit is used for inputting a pavement image to be detected into a pre-trained pavement disease segmentation model for segmentation and extraction to obtain a pavement disease region image of a pavement, wherein Unet networks comprising encoders and decoders are used as repeated units to be stacked to obtain a first generator, a first generation reactance network is formed according to the pre-trained first discriminator and the first generator, the first generation reactance network is trained, and when the training times reach the maximum times, the first generator with segmentation and extraction capacity for pavement diseases of the pavement image is obtained and is used as the pavement disease region segmentation model; The disease detection unit is used for inputting the disease area image into a pre-trained disease detection model to detect so as to obtain a disease detection result of the road surface, wherein the pre-built YOLOv neural network is trained, and the disease detection model is obtained when the first loss function converges.
10. An electronic device comprising a memory and a processor; A memory for storing a computer program, the computer program comprising program instructions; A processor for executing the program instructions to cause the electronic device to perform the steps of a method for detecting a road surface disease according to any one of claims 1 to 8.

Description

Road surface disease detection method, system and equipment Technical Field The invention relates to the technical field of highway pavement image processing, in particular to a highway pavement disease detection method, system and equipment. Background Roads are mainly classified into asphalt concrete roads and cement concrete roads, and the roads are classified into expressways, national roads, provincial roads, county rural roads and special roads (such as mining roads and scenic roads) according to functional grades. Asphalt concrete is subject to cracking (transverse, longitudinal and netlike), rutting, oil-flooding and other diseases caused by high-temperature softening, low-temperature shrinkage and vehicle heavy load, cement concrete road is subject to joint damage, reflection cracking, board angle breakage, peeling and other problems, road diseases (mainly comprising transverse cracking, longitudinal cracking, netlike cracking and the like) can not only shorten the service life of the road, but also directly influence the driving safety and comfort, especially road sections with large traffic flow such as expressways, national roads and the like, and the risk of traffic accidents caused by diseases is higher, so that rapid and accurate road disease detection is a key premise of road condition analysis and timely repair. In the related art, a pavement disease detection method based on YOLOv network model is widely tried, but is limited by a single frame-level target detection technical model, is difficult to adapt to complex disease scenes of different types of highways, and has the following problems: First, YOLOv network models cannot capture pixel-level features of road surface imperfections, including edge contours of fine cracks, interlaced textures of mesh cracks, internal hollows/depths of pits, pit ranges, etc., resulting in missing detail information. The detailed information is a core basis for judging the severity of the disease (such as whether the width of the crack exceeds a certain value and whether the pit depth reaches a maintenance threshold value) and calculating the quantitative index (area, length and volume) of the disease, and directly influences the scientificity of a maintenance scheme and the accuracy of engineering quantity accounting. Secondly, the rectangular frame YOLOv cannot be attached to irregular disease forms (such as bent longitudinal cracks and irregularly-shaped peeling areas), and a non-disease area (namely false detection) or an incomplete disease area (namely omission detection) is easily framed. For example, the actual coverage of the netlike cracks may exceed the detection frame, the banded concave areas of the ruts are difficult to be completely wrapped by a single rectangular frame, and the deviations can lead to misjudgment of subsequent maintenance decisions, so that the problems of excessive maintenance or incomplete maintenance occur. The road surface diseases have the characteristics of unbalanced types (such as small diseases such as flaking/mud pumping/oiling) and irregular shapes (such as reflection cracks and corner breakage), and for the disease characteristics (such as an oiling diffusion area and mud pumping exudation trace) without clear rectangular shapes, YOLOv cannot generate an effective detection frame, so that the condition of disease detection omission occurs. Disclosure of Invention The invention aims to provide a highway pavement disease detection method, system and equipment, which solve the problems that the detail information cannot be captured and the disease is missed when the pavement disease is detected in the prior art. The technical aim of the invention is realized by the following technical scheme: In a first aspect of the present invention, there is provided a method for detecting a disease of a road surface, the method comprising: Acquiring a road surface image to be detected of a road; Inputting a pavement image to be detected into a pre-trained disease area segmentation model for segmentation extraction to obtain a disease area image of a pavement, wherein Unet networks comprising encoders and decoders are used as repeated units to stack to obtain a first generator, a first generation reactance network is formed according to the pre-trained first discriminator and the first generator, the first generation reactance network is trained, and when the training times reach the maximum times, the first generator with segmentation extraction capacity for pavement diseases of the pavement image is obtained and is used as the disease area segmentation model; And inputting the disease area image into a pre-trained disease detection model to detect, thereby obtaining a disease detection result of the highway pavement, wherein the pre-built YOLOv neural network is trained, and the disease detection model is obtained when the first loss function converges. In one implementation scheme, the output part of the encoder is configured w