Research On Crack Detection Method Based On Deep Neural Network

The automatic detection of road cracks can help the road maintenance department find and repair the pavement damage as soon as possible,reduce the maintenance cost and improve the road safety.The automatic detection algorithm of image cracks is the current research hotspot.The traditional image detection methods are inefficient and vulnerable to environmental interference.The main work of this paper is to study how to improve the accuracy of pavement crack detection.Combined with deep learning technology,two sets of neural network models for crack detection are proposed,which are disassembled and compared on different data sets.Finally,a set of crack image recognition and detection system is developed based on the two network models.The specific work includes the following contents:(1)A bidirectional fusion network based on pyramid feature and cascade attention structure is proposed.The network uses resnext50 as the skeleton network to extract the pyramid features of road image,and then generates cross layer and cross scale two-level attention masks to enhance the convolution features of cracks.The two-way fusion method is used to generate the crack prediction map for the crack features of different scales,and the multi-stage crack prediction map is finally fused into the crack map output with the same size as the original input image.On the basis of resnet50,the network adds three new modules,namely layer attention(LA),scale attention(SA)and multi view enhance(MVE).The network has been trained and tested on cracktree260,crackls315 and Deep Crack DB data sets,and the ODS and OIS indicators have achieved high results.(2)A multi-scale fusion network based on encode and decode structure is proposed.In this network,Segnet is used as the backbone to extract road features,and skip connection is used to fuse the encoding layer and decoding layer.Finally,crack feature maps of different scales are fused to generate crack prediction map.In this network,lambda convolution module is used to replace 3 × 3 convolution to establish long-range feature dependence between pixels,and cross layer attention mask is generated at the same scale to enhance the fusion features at that scale.After experimental comparison,the network also performs very well in crack detection.In some data sets,the ODS and OIS indexes exceed the bidirectional fusion network of pyramid features and cascade attention structure.(3)A set of crack detection system is developed based on two kinds of network models.A set of crack detection system is developed by using image technology,packaging technology and user-defined file configuration technology.The system mainly includes image loading module,image information processing module,calculation and display module,parameter configuration module and data storage module.It can complete the operation of detection preview,crack custom evaluation,serious damage area summary report generation,etc.,to help the road maintenance department quickly and accurately locate the pavement damage.