A High Resolution Semantic Segmentation Method via Multi-branch Structure and Gating Mechanism

In order to solve the problem that HRNetv2 and other multi-branch structure networks cannot effectively fuse multi-level features in semantic segmentation tasks, a new multi-level feature fusion method based on the gating mechanism is proposed. Firstly, a gated fusion unit is constructed to fuse the feature information of multiple branches selectively. Secondly, a bottom-up fusion method is adopted to progressively enhance the feature representation of each branch by means of spreading semantically high-level features and detailed low-level features. Finally, features of branches are concatenated channel-wisely to output the predicted mask, and the bilinear interpolation algorithm is used to restore the original image size. Experiments show that the proposed method with only a few parameters achieves 77.01% mIoU and 80.43% mIoU in PASCAL VOC 2012 + Aug and Cityscapes respectively, increases by 1.14% mIoU and 1.92mIoU compared with HRNetv2-W48, and outperforms many baseline models.