Multi-scale Image Dehazing Network Based on Local Transformer

Most of the existing dehazing methods aim at hazy images with homogeneous haze, but the results obtained when dealing with hazy images with non-homogeneous haze are often unsatisfactory. In this paper, we propose an efficient local Transformer-based multi-scale image dehazing network (MIDNet), which takes advantage of the linear computing characteristics, the local information in the local window and the long-range relationship between pixels of the local Transformer to better handle hazy images with homogeneous and non-homogeneous haze. First, a multi-scale feature extractor based on local Transformer is designed to extract multi-scale features efficiently and comprehensively, making the network more generalizable and flexible. Secondly, the proposed feature aggregation module based on pyramid structure and the dense connection realizes the comprehensive aggregation of multi-source and multi-level features. Finally, in order to reconstruct the dehazed image with clear edges and rich details, the detail enhancement unit based on the gate structure is presented to retain the image more detailed information (such as edges, etc.). Extensive experiments on RESIDE, O-HAZE, I-HAZE, NH-HAZE and NITER datasets demonstrate that the proposed MIDNet achieves better dehazing performance.