Detail Guided Multilateral Segmentation Network for Real-Time Semantic Segmentation

Abstract

With the development of unmanned vehicles and other technologies, the technical demand for scene semantic segmentation is more and more intense. Semantic segmentation requires not only rich high-level semantic information, but also rich detail information to ensure the accuracy of the segmentation task. Using a multipath structure to process underlying and semantic information can improve efficiency while ensuring segmentation accuracy. In order to improve the segmentation accuracy and efficiency of some small and thin objects, a detail guided multilateral segmentation network is proposed. Firstly, in order to improve the segmentation accuracy and model efficiency, a trilateral parallel network structure is designed, including the context fusion path (CF-path), the detail information guidance path (DIG-path), and the semantic information supplement path (SIS-path). Secondly, in order to effectively fuse semantic information and detail information, a feature fusion module based on an attention mechanism is designed. Finally, experimental results on CamVid and Cityscapes datasets show that the proposed algorithm can effectively balance segmentation accuracy and inference speed.