MSSIF-Net: an efficient CNN automatic detection method for freight train images

Freight trains are one of the most important modes of transportation. The fault detection of freight train parts is crucial to ensure the safety of train operation. Given the low detection efficiency and accuracy of traditional train fault detection methods, a novel one-stage object detection method called the multi-scale spatial information fusion CNN network (MSSIF-Net) based on YOLOv4 is proposed in this study. The adaptive spatial feature fusion method and multi-scale channel attention mechanism are used to construct the multi-scale feature sharing network and consequently realize feature information sharing at different levels and promote detection accuracy. The mean average precision values of MSSIF-Net on the train image test set, PASCAL VOC 2007 test set, and surface defect detection dataset are 94.73%, 87.76%, and 75.54%, respectively, outperforming YOLOv4, Faster R-CNN, CenterNet, RetinaNet, and YOLOX-l. The detection speed of MSSIF-Net is 33.10 FPS, achieving a good balance between detection accuracy and speed. In addition, the MSSIF-Net performance is estimated after adding noise or rotating the train images at a slight angle to simulate a real scene. Experimental results indicate that MSSIF-Net has favorable anti-interference ability.

Automated summary

A novel one-stage object detection method based on YOLOv4, called the MSSIF-Net, is proposed for the fault detection of freight train parts. It achieves high detection accuracy and speed, outperforming other traditional methods. Furthermore, the MSSIF-Net demonstrates favorable anti-interference ability.