Deep Semantic Segmentation Neural Networks of Railway Scene

Abstract

It is crucial that the situation of railway scene needs to be periodically monitored to ensure railway safety. Usually, the camera is placed in the driver's cab on a moving vehicle to generate a large number of high-resolution images of railway scene. The traditional image processing inspection methods can only inspect railway scene of the local and simple background. They are feasible sometimes, but not universal and efficient. In this paper, a novel and practical approach is proposed to visual railway scene detection with Deep Semantic Segmentation Convolutional Neural Networks(DSSCNN). It is helpful for precisely locating and detailedly describe important parts of the railway scene in the entire image instead of separately detecting like traditional methods. In addition, three particular blocks named Conv Block, Dilated Conv Block, and Sum Block in DSSCNN are designed to extract more context information from railway images and enhance the segmentation performance. It shows that DSSCNN trained end-to-end and encoder-decoder network structure for semantic segmentation of railway scene are more superior than UNet and Fully Convolutional Networks(FCN) on the same datasets. At the same time, it is also verified that it is feasible and efficient to describe, visual and locate the key parts of the railway scene by DSSCNN.