부산대학교 도서관

Deraining of images plays a pivotal role in computer vision by addressing the challenges posed by rain, enhancing visibility, and refining image quality by eliminating rain streaks. Traditional methods often fall short of effectively handling intricate rain patterns, resulting in incomplete removal. In this paper, we propose an innovative deep learning-based deraining model leveraging a modified residual UNet and a multiscale attention-guided convolutional neural network module as a discriminator within a conditional generative adversarial network framework. The proposed approach introduces custom hyperparameters and a tailored loss function to facilitate the efficient removal of rain streaks from images. Evaluation on both synthetic and real-world datasets showcases superior performance, as indicated by improved image evaluation metrics such as PSNR, SSIM, and NIQE. The effectiveness of our model extends to improving both rainy and foggy images. We also conducted a comparative analysis of computational complexity in terms of running time, GFLOPs, and no. of parameters against other state-of-the-art methods to demonstrate our model’s superiority.