Crack Detection In Pavement Image Based On Deep Cross-Domain Generation

Crack detection in pavement images is of great significance to road scientific maintenance.At present,a lot of research results have been achieved in the detection of pavement cracks based on deep learning,but most of the methods based on deep learning require a large number of pixel-level labeled images to train the network model.However,obtaining a large number of pixel-level annotated images is time-consuming and expensive.In recent years,image depth generation technology has developed rapidly.In the field of image cross-domain generation research,in the process of cross-domain generation of target domain images from source domain images,through the constraints of specific loss functions,the generated target domain images retain certain features of source domain images.some specific attribute.Inspired by this,this thesis conducts research on new ideas and methods for pavement crack detection based on the basic principles of image cross-domain generation.The proposed method does not depend on pixel-level labeled images during the training process.The experimental results show the effectiveness of the proposed method,which has certain theoretical and practical significance for promoting the further promotion of pavement crack detection technology.The main work is as follows:1.A crack detection method for pavement images based on fixed-point generative adversarial networks is proposed.First,the fixed-point generative adversarial networks is used to convert the crack image into a crack-free image,and the difference image is obtained by subtracting the original image and the generated image pixel by pixel.The non-zero area in the difference image corresponds to the suspected crack area;Then,Le Net-5 is used to classify suspected crack areas into two categories,remove interference areas,and reduce the impact of road noise on the detection results;finally,use region growing to post-process the binary classification results to obtain the final detection results.The experimental results on the two public datasets of CFD and Deep Crack show that the method in this thesis not only achieves crack detection performance comparable to common deep learning methods such as Seg Net,U-Net,and FCN,but also requires only image-level annotations during the training process.Avoids reliance on pixel-level annotated images.2.In order to further improve the ability of fixed-point generative adversarial networks to extract crack features and promote the performance of crack detection,this thesis proposes a crack detection method based on mixed attention fixed-point generative adversarial networks.A mixed attention module composed of channel attention and spatial attention is added to the deep feature extraction module of the generator to improve the image conversion ability of the generator,so that the difference map can describe a more complete crack area.Compared with the fixed-point generative adversarial networks,the crack detection method based on the mixedattention fixed-point generation confrontation network not only inherits the advantages of the fixed-point generative adversarial networks not relying on pixel-level labeled images,but also improves the precision,recall rate and F1 score and other evaluation indicators.Both have been improved;compared with deep learning methods such as Seg Net,U-Net,and FCN,the detection performance has also been improved,indicating the effectiveness and necessity of adding a mixed attention module.