Lite-weight Skip Attention Network for Single Image Dehazing
Guoqing Dai, Bo Yang, Zhaozi Zu, Zhongjun Qu · 2025
Image dehazing is a low-level vision task aimed at recovering clear scenes covered by haze from hazy images, which is crucial for many high-level vision tasks such as object detection, semantic segmentation, etc. In this paper, we propose a lite-weight end-to-end CNN named LSA-Net (Lite-weight Skip Attention Network) to directly restore haze-free images from hazy inputs. In this work, we adopt multi-scale strategy to process hazy images, further extracting features through dilated convolutional residual blocks, using depth-wise separable fusion layer to fuse features from different levels, and incorporating an attention mechanism in the skip connection from input to output. The experimental results demonstrate that the proposed network can achieve good performance in terms of PSNR and SSIM, while maintaining a relatively low number of parameters and FLOPs, thereby keeping a good balance between performance and model complexity.