CycleGAN Image Defogging Method Based on Residual Dual Attention Mechanism

Abstract

The traditional image defogging method based on atmospheric scattering model is prone to color distortion and feature loss in the process of image defogging due to the variation of atmospheric scattering coefficient in the environment. This article focuses on the problem of color distortion and feature loss in image defogging caused by the CycleGAN network, We propose a CycleGAN image defogging method based on residual attention mechanism. Firstly, we add channel and spatial attention mechanisms to the residual network to form spatial and channel attention residual blocks, which are added to the two generators of CycleGAN to prevent color distortion during feature extraction. Secondly, we incorporate cyclic perceptual consistency loss, When the CycleGAN network learns images from two different style datasets, due to the fact that the two datasets being learned are clear images and foggy images, most foggy images are severely damaged. The cycle preserves the original image structure by looking at the combination of high-level and low-level features.