An Improved Faster R-CNN for Railway Fastening System Detection

Abstract

In the automatic railway anomaly inspection technology based on image processing and deep learning, an effective algorithm used for high-precision detection of the fastening system is very important, especially in turnout sections. It is challenging because the background of the turnout sections is complicated with various types of targets. This paper improved the Faster R-CNN model, used multi-scale feature map fusion for small targets. And modified predefined anchor to generate region proposals, added attention module to make the network focus on meaningful feature. Besides, this paper used cross-entropy function and SmoothL1 loss function for training and labeled 1200 image samples as dataset. Compared with the original Faster R-CNN model, the experimental results (AP) of the improved model in this paper increased from 96.3% to 98.9%, which effectively reduced the fault detection and missed detection and improved the accuracy of location.