Sleeper Defect Detection Based on Improved YOLO V3 Algorithm

In this paper, we present a detection method of sleeper defects based on the improved YOLO V3 algorithm to solve the shortcomings of the current track line maintenance which mainly is implemented by manual inspection, such as low efficiency and high risk factor. According to the characteristics of the background in the collected sleeper images, we optimize the weight value of the loss function. The K-means algorithm is used for clustering analysis of the sleeper data set, and then the optimal five sets of anchor box sizes are selected by the elbow method. For the purpose of improving the robustness of the model to different resolution images, we also adopt multi-scale training. The results of the experiment indicate that the improved YOLO V3 algorithm has obvious enhancement in the three performance indexes of Recall, Precision and mean Average Precision (mAP). Our work involving studies of intelligent identification of sleeper defects prove to be encouraging.