Single Image Dehazing with Global Detail Attention and Hierarchical Feature Fusion

Single image dehazing is a crucial task in computer vision that restores visual effects that approximate real scenes from images with low contrast and sparse detail features. Convolutional Neural Network are the mainstream technology in the research of single image dehazing. While, Transformer as an emerging approach, has not been widely applied in the research of image dehazing. This paper introduces improvements to the Vision Transformer, such as a Global Detail attention and a Hierarchical Feature fusion. Specifically, GDA encompasses Global Condensed Attention and Global Feature Enhancement, where GCA retains the deep interrelational information between entities, and GFE amplifies the global cross-dimensional interactions. HFF enhances model accuracy through cascaded convolutions without increasing computational complexity. Through numerous experiments, the proposed method has shown effective results in image dehazing. We obtained indices of 40.40 PSNR and 0.995 SSIM dataset on the indoor SOTS dataset, and both Param and FLOPs were smaller than the latest models.