Multiple Resolutions Detail Enhancement Network for Real-time Image Semantic Segmentation

Real-time image semantic segmentation draws the attentions of more and more researchers as a basis of scene understanding, and it has been applied in many fields that need fast interaction and response, such as autonomous driving and robot control. Considering the loss of low-level spatial information with the deepening network layer, we propose a multiple resolutions detail enhancement network (MRDENet) in this paper, which adequately extracts and utilizes accurate low-level detail information from original images with different resolutions. MRDENet consists of three light-weight branch sub-networks, and designs dense oblique connections between adjacent branches to preserve the different level effective features of previous branch. Furthermore, a new multi-level information aggregation module is presented to effectively fuse the low-level detail features and the high-level semantic features of different branches by employing group convolution and channel shuffle with low computation cost, thus ensuring that MRDENet could achieve a favorable trade-off between segmentation precision with inference speed. The experimental results show that MRDENet achieves 73.1% mIoU with 93 FPS on Cityscapes dataset, and 68.5% mIoU with 112 FPS on CamVid dataset, which indicates the performance of MRDENet is competitive with the state-of-art methods.