Multi-Scale Image Dehazing Network Based on Local Transformer

Aiming at the problem that most existing dehazing methods cannot deal with hazy images with non-homogeneous haze well, an efficient local Transformer-based multi-scale image dehazing network (MIDNet) is proposed. First, a multi-scale feature extractor is designed to efficiently and comprehensively extract multi-scale features, 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. Then a feature aggregation module is proposed based on the pyramid structure and dense connection to achieve comprehensive aggregation of multi-source and multi-level features. Finally, a detail enhancement unit based on the gated structure is designed to retain the image more edge and other detailed information. Extensive experiments on RESIDE, O-HAZE, I-HAZE, NH-HAZE and NITER datasets demonstrate that MIDNet has better visual effect. And on the NITER dataset, compared with the SRKTDN and DeHamer methods, the PSNR of MIDNet increases by 5.552 0 dB and 8.170 2 dB respectively, and the SSIM increases by 0.029 7 and 0.095 3 respectively.