Darkseg: a lightweight and edge-optimized network for nighttime semantic segmentation

Darkseg: a lightweight and edge-optimized network for nighttime semantic segmentation

연구 분야: Strategies

학회: Signal, Image and Video Processing

Semantic segmentation of nighttime road scenes is a critical task in autonomous driving. However, challenges such as insufficient lighting, over-exposure, and the scarcity of labeled data make this task particularly difficult. To narrow the gap between day and night domains, we introduce DarkSeg, a lightweight unsupervised domain adaptation network that offers an effective solution to these challenges. First, adversarial training techniques are utilized to achieve domain adaptation from the daytime dataset cityscapes to nighttime scenes, mitigating the issue of data scarcity. Second, the DarkSeg reduces inter-domain illumination differences through an enhanced low-light image restoration module (DarkNet), improving segmentation accuracy under low-light conditions. Third, the Polarized Self-Attention (PSA) module is integrated into the bottleneck module of the segmentation network, PASeg, to improve the representation of edge features under low-light conditions. Additionally, a boundary-based loss function (EdgeLoss) is introduced to improve the accuracy of edge pixels, thereby enhancing the overall clarity of the segmentation results. Experimental results show that DarkSeg performs significantly on the Dark Zurich and Nighttime Driving benchmark datasets, with mean Interaction over Union (mIoU) scores of 48.13 and 49.72 , respectively. This represents an improvement of 2.93 and 2.02 percentage points over the existing DANNet method. These results demonstrate DarkSeg’s superior performance in terms of network robustness and accuracy, which renders it especially suitable for semantic segmentation tasks at night.