An asphalt road crack image segmentation method based on a generative adversarial network

Abstract

The invention discloses an asphalt road crack image segmentation method based on a generative adversarial network. U-Net, CU-Net and FU-Net networks are used as a generator model G for generating an adversarial GAN models; GUNet, CUNet and FUNet contained in the generator model are respectively combined with three same binary classification network discriminator models D (Discriminative) to form UGAN, CUGAN and FUGAN models; mutual iteration competition optimization training is carried out through a generator and a discriminator, and finally trained generator model U-Net, CU-Net and FU-Net areused as crack image divider; U-Net, CU-Net and FU-Net models realize the image segmentation of the crack of the complex asphalt road. Compared with the prior art, the required training data set is less, the crack segmentation precision is higher, and the precision ratio and recall ratio are higher.

Description

technical field [0001] The invention relates to the technical field of image segmentation of asphalt road cracks, in particular to a method for segmenting cracks in asphalt road images using a deep learning generative confrontation network (GAN). Background technique [0002] Along with economic development, highway infrastructure construction plays an increasingly important role in the development of national economic construction. Regular maintenance and management of highways can reduce highway maintenance costs and traffic accidents. important parts of. Highway image crack segmentation is an important technology to assist highway monitoring and management automatic road surface disease detection. Many people have invested a lot of research on image crack segmentation technology through digital image processing technology. [0003] Due to the interference of random factors such as uneven illumination, random noise, blurred image gray levels, road signs and oil pollution ...

AI Technical Summary

Problems solved by technology

A common method is to divide the complete crack image into sub-images of equal size, and use a deep convolutional network (CNN) to classify the crack area and non-crack area of ​​the sub-image. Using CNN as a classification model can better extract image features. But the disadvantage is that a large amount of labeling data is required, and the accuracy of labeling data labeling is required to be high.

However, due to the reasons of the labeling personnel in the actual labeling, the blurred cracks in one image may be regarded as the cracked area, while the other image may be used as the normal area, which leads to the accuracy of the model when the data is blurred and the labeling is unclear. Higher rate, lower recall

Images