A Vision Transformer Dehazing Network Combining Multiscale Convolution and Attention

In recent years, convolutional neural networks have played a key role in dehazing, and as research has progressed, more and more forms of convolutional neural networks have emerged. However, most of these Mnetworks suffer from problems such as insufficient processing of the image's local details and inability to effectively focus on different scales and spaces, which leads to poor dehaze image results. To address these concerns, in this paper, we propose a dehazing network (DMCAFormer-t) that combines a new multiscale convolution and a feature fusion attention module to solve these problems and improve the effectiveness of image dehazing. In the DMCAFormer-t network, we design a new convolutional attention module (NCFB) based on the enhanced Swin Transformer structure to solve the effects of the transformer structure in dehazing, such as insufficient processing of local details. We designed the feature fusion attention module (FFAM) to solve the problem of inaccurate extraction of important features at different scales and in other spaces due to the traditional end-to-end dehazing network. Experimental results show that our proposed DMCAFormer-t network outperforms many algorithms in dehazing on the Reside dataset and achieves effective dehazing.