CESegNet:Context-Enhancement Semantic Segmentation Network Based on Transformer

CNN-based methods have achieved success in semantic segmentation. However, research on improving network robustness in this domain has been limited. Similarly, transformer and its variants have recently shown state-of-the-art results in many vision tasks, from image classification to dense prediction, because transformer has a global receptive field, but transformer-based methods have much higher computational complexity compared to CNN-based methods. To address these problems, we introduce a context-enhancement network based on the transformer. Firstly, we enhance global contextual information through a hierarchical simplified transformer encoder. Then, we design two different Context-Enhancement Modules (CEM) to enrich contextual features further. Finally, we propose a Contextual Fusion Decoder (CFD) to fuse multi-scale contextual information. Extensive experiments demonstrate that our method achieves significant performance and robustness compared to previous counterparts. Our best model, CESegNet-Large, achieves 82.21% mIoU and 48.77% mIoU on the Cityscapes and the ADE20K validation sets, and it also demonstrates excellent zero-shot robustness on the Cityscapes-N dataset.