Research On Recognition Technology Of Highway Pavement Crack Image Based On Deep Learning

Damage detection of highway pavement is an important part of highway maintenance and management.Quick and accurate access to highway pavement information is the key to highway maintenance management.At present,most of the detection methods of highway crack image are traditional image processing methods,but the traditional crack image processing methods are weak in adaptability and accuracy,and their automatic detection technology is not perfect.With the rapid development of computer hardware,deep learning technology is widely used in image recognition tasks.Based on the deep learning image processing technology,this dissertation takes the detection research of highway cracks image as the main line,and mainly works from the following aspects:(1)Aiming at the problems of uneven illumination,low image contrast,shadow of building and highway markings in highway pavement crack images,adaptive Gamma light correction algorithm,piecewise linear grayscale stretching algorithm,shadow removal algorithm based on K-Means clustering and adaptive gray-scale attenuation algorithm are used to preprocess the crack image,which lays the foundation for subsequent image segmentation and classification.(2)In order to extract the crack area of the highway crack image,an algorithm of highway crack image segmentation based on generative adversarial networks is proposed.The generative adversarial networks is composed of two sub-networks,a generator and a discriminator.The improved CU-Net and FU-Net networks based on U-Net are used to as generator network respectively,and then the binary classification network is used as the discriminator network.Through the combination of generator network and discriminator network,the performance of generator network is gradually enhanced.Finally,the generator network is used for road crack image segmentation.The experimental results show that the generative adversarial networks has higher image segmentation accuracy than the original algorithms.(3)Aiming at the problem of highway pavement crack images classification,a lightweight image classification network is proposed.This network uses the depthwise separable convolutions module to reduce the number of network parameters,and uses residual network jump layer connection to solve the problems of network gradient disappearance and network performance degradation.The experimental results show that the proposed network not only reduces the number of model parameters,but also ensures the classification accuracy and running speed of highway crack images.