Railway Image Semantic Segmentation Based on an Improved DeepLabV3+ Model

Scene understanding is a key prerequisite for intelligent railway systems and safe driving, with semantic segmentation being a critical component of scene understanding. The foundation for railway scene understanding is the accurate extraction of track regions. This paper addresses the challenge of feature extraction in complex railway scenes by proposing an improved DeepLabV3+ model that enhances the model's ability to discriminate and extract different types of railway infrastructure. Specifically, we introduce a lightweight, improved DeepLabV3+ model using MobileNetV2 as the backbone network. Additionally, we incorporate a Feature Enhancement Module (FEM) to capture rich local features, which improves the representation of small targets. We also replace the global average pooling layer in the Atrous Spatial Pyramid Pooling (ASPP) module with a strip pooling layer to better capture features of elongated and discrete regions. To fully leverage both shallow and deep features, we introduce the NAM mechanism to enhance the network's feature representation capabilities. Experimental results on the RailSem 19 dataset show that the proposed improved DeepLabV3+ model outperforms both the standard DeepLabV3+ model and other existing state-of-the-art models, significantly enhancing semantic segmentation performance in railway scenarios.