MixDehazeNet: Mix Structure Block For Image Dehazing Network

Image dehazing is a typical task in the low-level vision field. Previous studies verified the effectiveness of vanilla convolution kernel, transformer, and attention mechanism in dehazing. However, there are two main drawbacks in those methods: vanilla convolution and transformer have the short-comings of the insufficient receptive field and a large number of parameters respectively, and the previous design of the attention mechanism does not sufficiently consider an uneven hazy distribution. In this paper, a novel framework named Mix Structure Image Dehazing Network (MixDehazeNet) is proposed to solve the two issues mentioned above. Specifically, it mainly consists of two parts: the multi-scale parallel large convolution kernel module and the enhanced parallel attention module. Compared with a single vanilla kernel or transformer, parallel large kernels with multi-scale have a large receptive field and a relatively smaller amount of parameters, and the multi-scale characteristics of the image. It can restore a single pixel based on a large range of surrounding pixels and simultaneously recover texture details while capturing large hazed areas. In addition, an enhanced parallel attention module is designed according to atmospheric scattering models, which can extract shared global information and location-dependent local information of the original feature in parallel. It performs better at uneven hazy distribution. Extensive experiments on five benchmarks demonstrate the amazing effectiveness of our proposed methods. We achieved or approached state-of-the-art performance in five standard datasets. The code is released in https://github.com/AmeryXiong/MixDehazeNet.