A Multi-Feature Migration Fog Generation Model

The dehaze methods are limited because authenticity of the synthesized dataset has not yet met the requirements for dehazing in real-world scenarios. The traditional image domain migration methods exhibit an uneven problem in fog generation. We propose a characteristic transfer fog generative model (FogGAN) for the synthesis of haze datasets. Firstly, we propose a multi-feature fusion strategy for haze distribution based on the principle of atmospheric scattering. We use transmission maps, depth maps, and mask maps to obtain the distribution of haze and transfer the fused information to the source domain. Secondly, in order to improve the error fitting phenomenon of multicommon information in the target domain, we designed a multi-layer attention module (MAConv). It focuses the neural network on the features of fog and excludes interference from other content. To address the issue of missing details in generated images. We conducted experiments on the VOC2007 dataset. It demonstrates the effectiveness and the ability to improve existing dehaze methods.