EFRNet: Edge feature refinement network for real-time semantic segmentation of driving scenes

Abstract

In the semantic segmentation field, the dual-branch structure is a highly effective segmentation model. However, the frequent downsampling in the semantic branch reduces the accuracy of features expression with increasing network depth, resulting in suboptimal segmentation performance. To address the above issues, this paper proposes a real-time semantic segmentation network based on Edge Feature Refinement (Edge Feature Refinement Network, EFRNet). A dual-branch structure is used in the encoder. To enhance the accuracy of deep features expression in the network, an edge refinement module (ERM) is designed in the dual-branch interaction stage to refine the features of the two branches and improve segmentation accuracy. In the decoder, a Bilateral Channel Attention (BCA) module is designed, which is used to extract detailed information and semantic information of features at different levels of the network, and gradually restore small target features. To capture multi-scale context information, we introduce a Multi-scale Context Aggregation Module (MCAM), which efficiently integrates multi-scale information in a parallel manner. The proposed algorithm has experimented on Cityscapes and CamVid datasets, and reaches 78.8% mIoU and 79.6% mIoU, with speeds of 81FPS and 115FPS, respectively. Experimental results show that the proposed algorithm effectively improves segmentation performance while maintaining a high segmentation speed.