Research On Rail Surface Defect Detection Based On Improved YOLOv5s

With the rapid development of China’s rail transit,the total length of China’s railway operation has reached 155,000 km by December 2022,ranking first in the world.As the carrier of train operation,the rail is one of the most important infrastructures in the railway system.With the increase of the rail service time,the rail surface will produce irreversible damage.Once the damage occurs,it is easy to develop rapidly,which seriously threatens the safe operation of the train.Therefore,accurate and rapid detection of rail surface defects has become the key problem to be solved urgently for the development of rail transportation in our country.Therefore,in order to solve the problems existing in traditional detection methods and improve the detection accuracy and detection speed of rail surface defects,this paper studies the surface defect detection algorithm based on deep learning.The main work and innovation are as follows:1)Aiming at the problem of insufficient rail surface defect image samples,the methods of flip transformation,random cropping,brightness transformation and generative adversarial network are used to expand the rail surface defect image sample dataset.2)This paper proposes a rail surface defect detection method based on improved YOLOv5 s.Firstly,the CDConv convolution module was used to replace the original first layer Conv convolution module of YOLOv5 s backbone network,which improved the detection speed of the network.Secondly,in the tail of Backbone end and the tail of Neck end,the Swin transformer scheme is used to improve the C3 module of the original network,which can obtain better global information and richer context information.Then,GAM attention was introduced into PANET to make the network pay more attention to the features of dense damage areas,and a new detection head was formed to improve the detection accuracy of the network.Finally,the Soft-SIo UNMS loss function was used to replace the original CIo U,which accelerated the convergence speed of the algorithm and reduced the regression error.The experimental results show that the improved YOLOv5 s algorithm proposed in this paper has high detection accuracy and detection speed,and can realize the accurate and fast detection of rail surface defects.3)In order to improve the ability of portable devices to detect rail surface defects and meet the limitations of storage space and power consumption,based on the YOLOv5 s network architecture,this paper uses lightweight networks Mobile Net v3 and Shuffle Net v2 as the backbone network in YOLOv5 s respectively.The Shuffle Attention mechanism is introduced into the last layer of the backbone network,and two lightweight rail surface defect detection networks,Mobile-YOLOv5 s and Shuffle-YOLOv5 s,are constructed.The experimental results show that the proposed two lightweight networks greatly reduce the number of parameters and calculation of the network model under the premise of ensuring certain detection accuracy and detection speed.