Recognition Of Rail Surface Defects Based On Deep Learning Algorithm Research

With the rapid development of modern railroad technology,railroads occupy an indispensable position in people’s daily travel,the importance of regular safety inspection of railroad tracks is self-evident.With the rapid rise of artificial intelligence technology,a large number of neural network models have emerged,and different neural network algorithms meet different application scenarios.Therefore,it becomes the main research objective and content of this paper to explore the algorithm of fast and high accuracy and suitable for the intelligent identification of rail defects in rail defect detection.Based on deep learning technology,this paper proposes an intelligent recognition algorithm for rail surface defects based on deep learning.Firstly,we designed an image acquisition system to collect images of rail defects to provide the basis for the next study.Then,for the rail defect image in the acquisition process,because of light changes,texture background clutter and other interference,making the captured image more or less defective edge is not obvious,small size,background interference,low contrast and other problems.In this paper,two parts of digital image processing and deep learning are studied to improve the accuracy of rail defect recognition.In the digital image processing part,the acquired rail defect images are operated by projection method rail region extraction method,improved Retinex image enhancement algorithm,column mean method background modeling difference method and adaptive threshold segmentation processing method in turn to obtain the complete defect segmentation contour,so that it enhances the ability of subsequent network for defect feature extraction and localization.In the deep learning part,based on the advantages of simple structure of YOLOv4 network,mature network,strong scalability and fast detection,feature extraction and network structure are improved and optimized on the basis of YOLOv4.Firstly,the Res2 Net structure and CBAM attention mechanism are introduced to increase the perceptual field and defect location weights.Secondly,the bottomup PAN path enhancement structure is removed from the PANet structure to reduce parameter redundancy;then a new feature detection layer is added to improve the feature extraction capability for small targets;finally,the anchor frame is re-clustered using Kmeans++ clustering algorithm to reach full-size coverage of rail defects.Finally,the above algorithms are fused to obtain an algorithm model applicable to the intelligent identification of rail defects.Finally,in order to verify the effectiveness of the algorithm in this paper,comparative experiments are analyzed from four aspects: first,to design comparative experiments on the defect detection accuracy under different anchor frames,second,to design experiments on the ablation of the improvement strategies for both parts of image processing and deep learning,third,to design training and testing experiments between the algorithm in this paper and other target detection algorithms on this dataset,and fourth,to design comparative experiments between this algorithm and the research results of other researchers on the publicly available small target dataset RSDDs.The experimental results surface that the average accuracy of the four defects of rail defect detection reaches 92.68%,the recall rate reaches 92.33%,and the average detection time of each image reaches 0.068 s,which is significantly better than other comparative algorithms.The experimental results prove that the algorithm in this paper achieves good results in detection performance and in real-time,and is more suitable for performing the task of rail surface injury detection.