Single Image Dehazing Based on Dynamic Convolution and Transformer

In this paper, an end-to-end multi-stage dehazing network based on convolution and Transformer is proposed. The network design is divided into three parts: encoding network, feature fusion network and decoding network. The encoding network extracts primary features of the haze image, the feature fusion network uses the serial Transformer module and the dynamic convolution module to make the information extracted from the features richer, and the decoding network is used to recover the image resolution. In the resolution restoration stage of the decoding module, the MixUp module is used to restore the high resolution image by combining the extracted primary features to reduce the loss of information. Extensive experiments were conducted on synthetic and real datasets to validate the role of Transformer module and Dynamic Convolution module in dehazing respectively. The results show that the proposed method achieves a good objective evaluation score and reconstructs a subjectively better dehazing image with a PSNR of 35.37 and a SSIM of 0.9849 on the SOTS test dataset.