Development of a Modified CycleGAN Model with Residual Blocks and Perceptual Loss for Image Dehazing

Sani Moch Sopian (Department of Electrical Engineering, Faculty of Engineering, Garut University, West Java, Indonesia)
Arief Suryadi Satyawan (Research Center for Telecommunication – BRIN, Indonesia)
Mokhammad Mirza Etnisa Haqiqi (Department of Electrical Engineering, Faculty of Engineering, Garut University, West Java, Indonesia)
Helfy Susilawati (Department of Electrical Engineering, Faculty of Engineering, Garut University, West Java, Indonesia)
Beni Wijaya (Department of Electrical Engineering, Faculty of Engineering, Garut University, West Java, Indonesia)
Khaulyca Arva Artemysia (Department of Electrical Engineering, Faculty of Engineering, Garut University, West Java, Indonesia)
Firman (Department of Electrical Engineering, Faculty of Engineering, Garut University, West Java, Indonesia)
Muhammad Ikbal Samie (Department of Electrical Engineering, Faculty of Engineering, Garut University, West Java, Indonesia)

Teknika
Vol. 14 No. 2 (2025): July 2025

Fog reduces image contrast and clarity, creating challenges for applications such as autonomous driving and remote sensing. This study proposes a series of CycleGAN modifications for single image dehazing using unpaired data, integrating residual blocks, attention mechanisms, VGG19-based perceptual loss, and haze-aware loss. Among ten architectural variants, Modification 10 combining perceptual and haze-aware loss achieved the best overall performance. Quantitatively, it showed stable generator losses (0.91 for Gen G, 0.57 for Gen F), with improved discriminator performance (Disc X: 0.59, Disc Y: 0.47), indicating better training stability and image realism. Additionally, it offered competitive PSNR (7.99), strong SSIM (0.4202), and low LPIPS (0.6577), confirming its effectiveness in both pixel-level accuracy and perceptual quality. Qualitatively, this model generated clearer, more natural images with improved edge sharpness and detail preservation. These findings demonstrate that the modified CycleGAN significantly enhances dehazing performance and presents a valuable contribution to deep learning-based image restoration.