Abstract: Currently, deep learning-based image dehazing algorithms struggle to simultaneously extract the global and local features of images, resulting in the loss of detailed information in the restored images and the occurrence of color distortion. To address this issue, a pixel-wise detail compensation dehazing network based on Transformer is proposed, which mainly consists of a Transformer-based encoder-decoder and a CNN branch. When a foggy image is input, global feature extraction is performed through the encoder. The Transformer in the encoder is composed of a channel attention block (CAB), a compression attention neural  block (CANB), and a dual-branch adaptive neural block (DANB). The CANB captures the global dependencies of image superpixels through feature aggregation, attention calculation, and feature restoration. The DANB adopts a dual-branch structure to encapsulate the global dependencies of superpixels into individual pixels, thereby obtaining global feature information. Meanwhile, the spatial attention in the CNN branch can enhance the model’s ability to perceive different fog densities and perform local feature extraction. Finally, in the decoder part, the features extracted by the encoder and the CNN branch are fused to output a clear image. Experimental results show that the proposed model performs excellently on both synthetic dataset (RESIDE) and real datasets (O-HAZE and NH-HAZE), and can effectively solve the problems of detailed feature loss and color distortion.

Key words: image dehazing, deep learning, dual branch feature fusion, detail compensation, Transformer architecture