Enhancing Hazy Image Dehazing Using a Modified CycleGAN

The degradation of image quality due to haze poses a significant challenge in various visual applications such as autonomous vehicles, surveillance, and remote sensing. This study proposes enhancements to the Cycle-Consistent Generative Adversarial Network (CycleGAN) architecture through multiple modifications aimed at improving dehazing performance on unpaired images. Ten variations of the generator architecture were developed using combinations of residual blocks, dilated convolutions, self-attention mechanisms, and the integration of additional loss functions such as perceptual loss and haze-aware loss. The dataset used consists of 50 hazy images and 50 clear images, trained under an unsupervised learning scheme. Evaluation was conducted quantitatively using PSNR, SSIM, and LPIPS metrics, as well as qualitatively through visual output analysis. The results show that Modification 5, which utilizes eight residual blocks, achieved the best performance with the highest PSNR (8.20) and lowest LPIPS (0.6443), while Modification 3 recorded the highest SSIM score (0.4419). This study demonstrates that internal architectural modifications to CycleGAN can effectively enhance dehazing image quality, even when trained on a limited and unpaired dataset. For future development, it is recommended to test the model on standard benchmark datasets and conduct direct comparisons with other dehazing models to strengthen the scientific contribution.