Research On Pavement Crack Image Detection Method Based On Self-supervised Deep Contrastive Learning

With the continuous development of traffic infrastructure in our country,the traffic mileage of the road increase year by year.Different kinds of injuries are mainly represented by the crack.Pavement cracks will not only affect the service life of the road,but also may endanger the driving safety,so the detection of pavement cracks is particularly important.The traditional detection methods cannot deal with the cracks in time and effectively,and it is more and more difficult to meet the actual needs.At present,the research of pavement crack image detection based on computer vision has become the mainstream,including crack image classification and crack image segmentation.However,the current research on pavement crack image based on computer vision generally has the defects of over-reliance on annotated information and insufficient processing of fine cracks.Self-supervised deep contrastive learning has the advantage of not needing annotated information and has many applications in image classification and image segmentation.Therefore,this thesis proposes the pavement crack image classification method based on deep contrastive learning and pavement crack image segmentation method based on deep contrastive learning.The main work is as follows:1.The investigation and research of pavement crack image detection methods based on deep learning are carried out,and the shortcomings of current pavement crack image detection methods based on deep learning are summarized.The deep contrastive learning model is proposed to process pavement crack images,and the investigation and research of deep contrastive learning in image classification and segmentation are carried out.The characteristics of pavement crack image are summarized.2.An image classification method of pavement crack based on deep contrastive learning is proposed.By using unmarked pavement crack image data and deep contrastive learning,the self-supervised accurate classification of transverse crack,longitudinal crack and mesh crack is realized.Furthermore,multi-scale features are integrated in the feature extraction part of deep contrastive learning,and the classification accuracy rate of 92.1% is achieved in the Crack500 dataset,which significantly improves the feature extraction ability of fine cracks and improves the noise robustness of the algorithm.3.An image segmentation method of pavement crack based on deep contrastive learning is proposed,which combines deep contrastive learning with improved U-Net structure.The method includes two stages: pre-training and fine-tuning.In the pre-training stage,the encoder part of the improved U-Net network learns the potential feature representation from the unmarked crack image.The training data uses the crack region image and the pavement background image,so that the model learns the distinguishable mapping relationship between the crack region and the background region.In the fine-tuning stage,the network loads the parameters of the encoder after the pre-training,and uses the marked training data for the retraining.Experimental results show that the proposed method achieves 91.7% segmentation accuracy in CrackForest and Crack500 data sets without increasing the number of existing training samples and their labeling,which significantly improves the model’s ability to distinguish background from cracks in pavement crack images,and improves the segmentation effect of cracks.