New Depth-Wise Asymmetric Bottleneck Network with Multi-Scales for Real-Time Semantic Segmentation

Real-time semantic segmentation is currently a hot research point in computer vision. In this paper, we focus on improving the performance of the network by modifying the architecture of DABNet. In this paper, we propose a new Depth-wise Asymmetric Bottleneck Network with Multi-Scales for Real-time Semantic Segmentation (MS-DABNet) that fuses multi-scale information to enhance the effectiveness of the network mainly by capturing and fusing multi-scale information to enhance the effectiveness of the network. Specifically, the original DAB module is improved to a self-correcting DAB module with adaptive acquisition of multiscale information, and the Cascaded and Cross Stage Partial ASPP (CCSP-ASPP) module proposed in this paper is introduced into the DABNet to obtain a superior MS-DABNet. The CCSP-ASPP module proposed in this paper is based on the improvement of the ASPP module, which utilizes the cross-stage feature fusion strategy idea and bottleneck structure, which greatly reduces the amount of computation while broadening the network structure so that each network layer can learn richer features. The proposed MS-DABNet improves the mIoU from 70.10% to 72.84%, and achieves a running speed of 38.6 FPS to process high-resolution images on a single RTX 3090 GPU with a model size of only 1.33M.