Raindrop Removal On A Single Image Using The Generative Adversarial Network

Abstract

The presence of raindrops attached to the window glass or vehicle glass reduces visibility of the actual scene. The area covered by raindrops forms a round image and reflects light from the actual scene, this image is called Raindrop. In some cases, the camera's focus is blocked by Raindrop will result in a blurry image. The problem faced is that there is no actual landscape image, so to overcome this this research tries to adapt the research from Rui Qian and MaybeShewill-CV which uses the Generative Adversarial Network architecture, by adding the Raindrop and Groundtruh datasets from observations at Darussalam Gontor University. The purpose of this study is to remove raindrops from a single image. This is important to research because it provides updates and optimizes the results of previous research. This study shows the accuracy of PSNR 21.37 and SSIM 0.7561. The model managed to remove Raindrops from the image, but still couldn't match the Groundtruth image. Inability to handle raindrops due to lack of time to run a large number of epochs to produce PSNR values ​​above 40 db and SSIM above 0.9. PSNR and SSIM values ​​can continue to increase along with the addition of the dataset as well as the number of epoch training models carried out.