Research On Road Crack Detection Method Based On Deep Learning

Pavement crack is one of the important manifestations of road diseases,as well as the early forms of many other diseases.China invests a lot of manpower and material resources in crack detection and maintenance every year.Automatic detection of pavement cracks can greatly reduce maintenance costs.At present,the detection methods using laser,radar and other related technologies are relatively mature,but due to the expensive equipment,they are only suitable for special roads such as bridges,and it is difficult to apply them on a large scale.With the rapid development of computer vision,road crack detection based on deep learning has also achieved certain results,but in practical applications,there is still room for improvement in detection performance.In this paper,the following researches are carried out for the automatic detection of road cracks:First of all,this paper investigates the research history and current situation of pavement crack detection,classifies the existing crack detection methods,analyzes the problems existing in the actual use of the existing crack detection algorithms,and introduces in detail the methods used in this paper related technologies include convolutional neural networks,recurrent neural networks,deep supervised learning,and multi-scale feature fusion.Secondly,this paper proposes a road crack detection method based on multi-stage structure feature extraction.Existing road crack detection methods based on deep learning are easily affected by changing and complex road conditions,have poor robustness,and require a large number of accurate markings.The method proposed in this paper considers the morphology and structure of cracks from a macro perspective,extracts the structural features of the cracks in a multi-stage progressive manner,and merges them with local features for crack detection.This paper also uses deep supervision for training to visualize the detection results at each stage to ensure the validity and transparency of the structural features obtained at each stage.The experimental results show that the road crack detection method based on multi-stage structure feature extraction can obtain better detection results than existing methods.Finally,a pavement crack detection method based on multi-stage continuous feature extraction is proposed.Considering that most of the existing forms of cracks are transverse or longitudinal,the structural features of cracks obtained by convolution method introduce a lot of irrelevant information.According to the connection relationship between fracture image blocks,a convolution structure of LSTM2 D module is designed to replace the multi-stage structural feature extraction method.In this paper,the LSTM2 D module is used to get the continuity characteristics of cracks by introducing LSTM,and the deep supervision method is also used for training.The experimental results show that the pavement crack detection method based on multistage continuous feature extraction can achieve better results than the pavement crack detection method based on multi-stage structural feature extraction with less parameters.