Real-time semantic segmentation network based on improved BiSeNet V1

Abstract

Aiming at the problems of unclear boundaries and low segmentation accuracy in general real-time semantic segmentation networks, a real-time semantic segmentation network based on improved BiSeNet V1 was proposed. Based on the BiSeNetV1 network, a spatial enhancement module (SRM) is introduced into the spatial path to enhance the spatial information and improve the detection ability of target boundaries and small targets. At the same time, when the spatial path and context path feature information are fused, the Feature Aggregation Module (FAM) is proposed to solve the difference in feature representation between the two paths in feature fusion and improve the fusion efficiency. We experiment on Cityscapes that reach 69.6% MIoU and 100.4 FPS. The experimental results show that the proposed algorithm can improve the efficiency of segmentation.