Polarization image defogging based on detail recovery generative adversarial network

Abstract

Most of the existing polarization image based defogging methods use a priori and assumptions to recover images, and although these methods have made great progress, the a priori or assumptions are not always reliable in practical scenarios, which limits the performance of defogging methods. In this paper, we design a two branch network to learn image features; the defogging network uses the fusion module to adaptively assign weights to different path features to ensure the defogging effect while focusing on the haze region; the detail recovery network uses convolutional layers and smoothly expanding convolution to expand the perceptual field and fully obtain local and global feature information; finally, the features obtained from the detail recovery network and the defogging network are fused to improve the defogging Finally, the features obtained from detail recovery network and defogging network are fused to improve the defogging effect. The experimental results show that this method can reconstruct clear images in foggy environment with 2db improvement in PSNR and 0.016 improvement in SSIM, and the quality of reconstructed images is better than existing algorithms.