| Road transport is one of the two basic ways of transportation which is not only an independent transportation system but also an important way of material distribution.Therefore,how to guarantee the safety of the road transport has been a hot topic.Road surface cracks are a very common defect that seriously affect the service life of the road and threaten the safety of road transport.Therefore,timely and accurate automated crack detection is critical to road maintenance.In this paper,we propose a new improvement scheme in view of the lack of precision and convenience of the existing crack detection methods.The main work of this paper is as follows:(1)In view of the current crack detection networks,the receptive field is relatively single and the context information of different scales is not fully utilized.Therefore,we propose a multi-scale crack detection network,MSC-Net.The network exploits U-Net as the basic structure,and introduces a multi-scale crack module(MSCM)into it,which enriches the receptive fields of the network at different scales without losing position information as much as possible.Multi-stage fusion supervision(MSFS)is introduced into the network to supervise the detection results of fusion features at different stages.And the edge regularization term is introduced into the network to improve the positioning of crack edges.Because there is no recognized dataset for crack detection,and the crack size and background texture of existing crack datasets are single,we build a road crack dataset(RCD)which contains cracks of various sizes and types,as well as various background textures.The dataset contains 350 crack images,including horizontal,vertical,and meshed cracks,and six different types of background textures,which ensures the diversity of cracks and the difficulty of crack detection tasks.The experimental results on the RCD crack dataset in this paper show that MSC-Net has achieved a 6.8%improvement in F1-score compared to Deep Crack,which is currently performing better in this field,and has improved detection precision and recall 8.35% and 4.41%.The results prove the superiority of the multi-scale crack detection network MSC-Net,which significantly improves the crack detection effect.(2)A semi-supervised crack detection method based on fusion strategy is proposed.In the crack detection methods based on image semantic segmentation,pixel-level image labeling is time-consuming and laborious,and the accuracy is difficult to guarantee,which greatly hinders the application in practical tasks.In order to solve this problem,we propose a simple and effective semi-supervised method,which uses a variety of unsupervised algorithms to fuse its prediction results as pseudo-labels,thereby simplifying the generation of pixel-level pseudo-labels.This method only needs a small amount of fully supervised pixel-level labels and uses a large number of pseudo-labels,which can greatly improve the crack detection precision of the network.The experimental results on the RCD crack dataset show that under the premise that the ratio of the number of images in the fully-supervised dataset and the unsupervised dataset is 1/15,the method achieves the same experimental effects as SegNet and DeepCrack under fully-supervised conditions.This result proves the effectiveness of the semi-supervised crack detection method,which significantly reduces the dependence on manual labeling information and the burden of manual labeling under the premise of ensuring a certain detection effect. |
