Lunet: an enhanced upsampling fusion network with efficient self-attention for semantic segmentation

Semantic segmentation is an essential aspect of many computer vision tasks. Self-attention (SA)-based deep learning methods have shown impressive results in semantic segmentation by capturing long-range dependencies and contextual information. However, the standard SA module has high computational complexity, which limits its use in resource-constrained scenarios. This paper proposes a novel LUNet to improve semantic segmentation performance while addressing the computational challenges of SA. The lightweight self-attention plus (LSA++) module is introduced as a lightweight and efficient variant of the SA module. LSA++ uses compact feature representation and local position embedding to significantly reduce computational complexity while surpassing the accuracy of the standard SA module. Furthermore, to address the loss of edge details during decoding, we propose the enhanced upsampling fusion module (EUP-FM). This module comprises an enhanced upsampling module and a semantic vector-guided fusion mechanism. EUP-FM effectively recovers edge information and improves the precision of the segmentation map. Comprehensive experiments on PASCAL VOC 2012, Cityscapes, COCO, and SegPC 2021 demonstrate that LUNet outperforms all compared methods. It achieves superior runtime performance and accurate segmentation with excellent model generalization ability. The code is available at https://github.com/hbzhou530/LUNet.