WFAGAN-Dehaze: Wavelet-Enhanced Feature Attention CycleGAN for Single Image Dehazing
Yang Liu, Zhi-Lin Fan, Fei Wang, Yulong Wang, Dajun Du · 2025
Hazy weather severely impacts marine operations like autonomous driving and security monitoring. Images captured under hazy conditions frequently exhibit color distortion and loss of details. Moreover, it is difficult to acquire paired hazy-clear images with accurate pixel alignment. To solve these issues, a CycleGAN-based dehazing model for unpaired single images is proposed. The generator includes a Wavelet-enhanced Feature Attention (WFA) Module, which uses wavelet transforms to improve details and edge restoration. It also applies multiple attention mechanisms to focus on specific regions. In addition, the Multi-Scale Dilated Convolution (MSDC) Module captures features at different scales, resulting in images with more clarity and details. Quantitative metrics such as PSNR and SSIM demonstrate the effectiveness of the proposed model on benchmark datasets.