BRNet: A Bilateral Residual Network for Real-Time Semantic Segmentation

The lightweight semantic segmentation network makes use of various excellent residual designs, and these residual modules not only improve the efficiency of semantic segmentation but also enhance the accuracy of semantic segmentation. Inspired by the structure of residuals, we propose a lightweight network, based on the modified BiSeNet V2 model, termed the Bilateral Residual Network (BRNet). Additionally, the Three Branches Residual Stem (TBR-Stem) module is also designed. Specifically, the two-branch structure of the original Stem Block is modified to a three-branch structure, and the asymmetric convolution operation is adopted, while depth-wise separable convolution and dilated convolution are introduced. As a result, TBR-Stem significantly reduces the computation, expands the receptive field, and enhances the feature extraction capability. Besides, the Channel Attention Module (CAM) is added to the Semantic Branch to reduce channel noise. The method achieves 73.3% MIoU on the CityScapes validation set and 143 FPS inference speed on a single NVIDIA GeForce GTX 1080Ti, finally, demonstrating that the method improves segmentation accuracy.