SIDC-GAN: Single Image De-Rain Cascade UNet GAN

Single image de-raining focuses on removing rain effects from an image while preserving important details, making it one of the challenging tasks in computer vision. This article introduces the Single Image De-Rain Cascade-UNet Generative Adversarial Network (SIDC-GAN). The framework consists of a lightweight UNet-based generator and a discriminator working adversarially to enhance the de-raining process. The generator, a lightweight UNet structure that consists of a convolution block, inverted residual block, and inverted residual block with squeeze connection, aims to produce realistic, rain-free images, while the discriminator differentiates between real and generated images. The inverted residual block utilizes depth- wise convolutions to significantly reduce computational cost compared to regular convolution, thereby enabling a lightweight UNet-based generator architecture. The adversarial interaction drives continuous improvement, resulting in higher-quality, de-rained outputs. SIDC-GAN, trained on the extensive Rain13 K dataset, demonstrates robust performance in diverse and complex rainy scenarios due to its exposure to a wide range of rain conditions. Evaluation of five benchmark data sets—Rain100 L, Rain100H, Test100, Rain1200, and Test2800—shows that SIDC-GAN achieves significant improvements in terms of peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). The SIDC-GAN model shows a 2.6% increase in PSNR compared to state-of-the-art methods. The SIDC-GAN model further evaluates the effect of cascading multiple UNet networks to progressively refine the results in the generator. The result shows that two consecutive UNet layers have achieved an improvement of 3.45% in PSNR and 4.066% in SSIM compared to a single UNet layer in the generator. However, increasing the number of UNet layers in the generator from two to four observed a diminishing return in output.