Two‐stage single image dehazing network using swin‐transformer

Hazy images often have color distortion, blur and other visible visual quality degradation, affecting the performance of some advanced tasks. Therefore, single image dehazing has always been a challenging significant problem. Convolutional neural network widely used in task, but limitations convolutional operation limit development task. Nowadays, Transformer offers holistic approach to CV does not grow location as deepens. For this reason, hierarchical is introduced for use network. Specifically, codec improved CNN are combined achieve basic feature extraction first stage. The encoder only models global relationship at each layer, reducing resolution map continuously expanding field perception. In addition, an inter-block supervision mechanism added between unit decoder refine features supervise select them, thus improving efficiency transmission. second stage, original block extract local features, then fusion interaction carried out. ensure authenticity transmission characteristic signals stage improve network, attention stages. It adds residual early input acquired passes next Ablation experiments show that two-stage benefits effects. experimental results on RESIDE, O-Haze, I-Haze datasets method superior methods effectiveness.