Removing Rain from Single Image Using Atrous U-Net and GAN

Unfavorable weather circumstances, characterized by heavy rainfall, not only reduce image visibility, however it also generates challenges for the efficiency of computer vision algorithms. This study endeavors to address the challenge of eliminating rain from single images, a process known as image de-raining. It specifically targets situations marked by dense rain streaks and the accumulation of rain, resembling mist or fog. This is achieved by leveraging the spatial distribution of rain streaks, referred to as the rain map. CGANs are employed for single-image de-raining, ensuring that the resultant derained images closely approximate their ground truth counterparts. The generator network utilizes the Atrous convolution-based U-Net architecture in which dilation filters are employed as kernels. The discriminator network employs the Patch Generative Adversarial Network (PatchGAN) for enhanced performance by discriminating original derained image and fake derained image. The suggested methodology highlights attaining robust numerical, visualized, and distinguishing performance in its objective function. Experimental results, verified on synthetic and real images, demonstrate the superior performance of proposed approach compared to recent cutting-edge single image de-raining techniques in terms of both numerical measures and visual quality. This research offers valuable insights into the integration of the U-Net architecture in deraining model. This superiority of the model is evident accross the evaluation metrics i.e., Peak Signal-to-Noise Ratio (PSNR) of 26.73 dB, Structural Similarity Index Metrics (SSIM) of 0.896, Universal Image Quality Index (UQI) of 0.698 and Visual Information Fidelity (VIF) of 0.5719 for single-image deraining, providing a robust solution for real sworld applications in computer vision.