Improving Nighttime Semantic Segmentation with Dual-Domain Mixed Attention and Shallow Feature Enhancement

Semantic segmentation in nighttime scenes has garnered considerable attention in the field of computer vision. Recent studies have begun to utilize frequency information to tackle this task. However, existing methods based on frequency solely consider the unidirectional guidance of frequency domain information on spatial features, disregarding the impact of spatial features on the frequency domain information. Moreover, the color and contour information of nighttime images is not prominent, leading to an insufficient contribution of shallow features in nighttime semantic segmentation networks. To address these issues, we propose a dual-domain mixed attention (DDMA) module and a shallow feature enhancement (SFE) module for nighttime semantic segmentation. Specifically, the DDMA module performs a domain transformation on the saliency learned from spatial and frequency domain features, enabling the cross-domain fusion of information in both frequency and spatial domains. The SFE module is employed to enhance the expressive capacity of shallow features within semantic segmentation networks. Extensive experiments on the NightCity, ACDC-Night and BDD100k-Night datasets showed effective improvement in nighttime semantic segmentation performance compared to methods using the same baseline, achieving state-of-the-art results.