Research On Real-time Detection Algorithm Of Rail-surface Defects Based On Improved YOLOX

Abstract

Real-time detection of rail surface defects is an important part of the future railway automation detection. Because of rail surface defects show multi-scale, small difference between background and prospect, therefore, a improved YOLOX real-time defect detection algorithm is proposed. concretely: Firstly, by counting the size of the defects, we design a backbone feature network which is suitable for detecting the multi-scale defects. secondly, in feature fusion network, a multi-scale network of Feature Pyramid Network and Path Aggregation Network is designed to extract richer semantic information and achieve more accurate spatial information. Thirdly, 9 Convolution Block Attention Modules are added between backbone network and feature fusion network as well as in the up-sampling process of feature fusion network, the purpose is to assist model training and update weights of model. Besides, replacing the Spatial Pyramid Pooling Module in the original model with the Atrous Spatial Pyramid Pooling Model, which can speed up the model training and get better training effect. Finally, we modify the loss function—to train the model with Focal-Efficient Intersection Over Union function in the regression loss function, it can get more accurate regression boxes and solve the problem of positive and negative samples imbalance. Ultimately, the experiment shows that the resulting algorithm can achieve 96.1% accuracy and 49.74 fps value, which is a relatively accurate and more reasonable real-time rail surface defect detection algorithm.

Keywords

• rail surface defect
• yolox
• real-time
• convolution block attention module
• atrous spatial pyramid pooling
• focal-efficient intersection over union