BSNet: A bilateral real-time semantic segmentation network based on multi-scale receptive fields

Abstract

The rise of autonomous driving and mobile robots has drawn attention to real-time road scene segmentation. However, category confusion, incomplete segmentation and inaccurate detail contours are common issues encountered in road traffic scene segmentation tasks. To tackle these road segmentation challenges, we proposes a bilateral real-time semantic segmentation network based on multi-scale receptive fields(BSNet). To enhance the segmentation accuracy of detail contours, an auxiliary detail detection module is introduced in the spatial branch. In the semantic branch, strip pooling and channel attention are added to the short-term dense concatenate module to enrich multi-scale receptive fields and strengthen channel features, addressing issues such as incomplete segmentation and category confusion. The final design of the complementary dual-guided fusion module improves feature fusion effectiveness. Experimental results on two benchmark datasets demonstrate that BSNet significantly improves these road segmentation difficulties and achieves competitive results in both speed and accuracy compared to the current state-of-the-art methods.