The Research Of Pavement Crack Recognition Algorithm Based On Deep Learning Multi-model Cascade

As one of the most common road diseases,pavement cracks are the most difficult points in the road maintenance process.Traditional manual detection pavement cracks methods have low efficiency and poor safety.In response to the above problems,this thesis uses the crack image data collected by the pavement detection vehicle to propose an intelligent recognition method that integrates road crack classification,detection and segmentation.This thesis main research contents are as follows:Firstly,this thesis analyzes the characteristics of the crack images collected by the road detection vehicle.At the same time,this thesis makes detection and segmentation annotations on the crack images and the annotates data is divided into training set,test set,and verification set.The Faster-rcnn and Yolov4 two classic target detection methods are used for model training and parameter tuning.On this basis,the feature extraction network of Faster-rcnn is improved by using the residual network.The detection results of Faster-rcnn,improved Faster-rcnn and Yolov4 are compared from the detection index,training and test time,and detection effect images.The results show that the Faster-rcnn-Res Net152 improves the accuracy of pavement crack recognition,and the average accuracy of pavement crack classification and detection reaches 90.42%.Secondly,this thesis uses Deep Labv3+ and U-Net two classic image segmentation networks to segment road cracks through model training and parameter tuning.On this basis,the U-Net network is improved using the attention mechanism.The segmentation effect of Attention-U-Net,U-Net and DeepLabv3+ are compared and analyzed.The results show that the Attention-U-Net network with attention mechanism has a better segmentation effect on pavement cracks,and the IOU index of pavement crack segmentation reaches 86.3%.Finally,this thesis proposes a road crack recognition algorithm based on multi-model cascaded of deep convolutional network,which cascades the road crack detection model and segmentation model.And this thesis uses the edge connection growth algorithm to optimize the road crack detection and segmentation results.At the same time,the geometric parameters of the crack are calculated according to the binary image of the crack output by the segmentation model.And the recognition results of pavement cracks are visually displayed on the original crack images.Experiments show that the multi-model cascaded road can accurately and efficiently identify the crack type and location of road cracks.And the crack recognition algorithm can calculate the geometric parameters of road cracks,which provides an important reference for road maintenance in the later stage.