Conditional Generative Adversarial Networks를 이용한 물방울 왜곡 복구

Abstract

Due to the recent development of artificial intelligence technology, self-driving cars and smart cars are drawing attention. These technologies get many problems from development and application, and the biggest one is safety. In case of self-driving cars or smart cars, camera lenses are used to recognize obstacles and objects around them. Current technologies are used for driving safety purposes, other uses of camera lenses are preventing crime, but if they start with CCTV (Closed-circuit televisional), they can also cause a variety of signals and speed controls, as well as station-stop violations.

The restoration of a polka-down distorted image is divided into the detection of water droplets and the image restoration. Using the properties of water droplets to detect them and then replace the detected area or restore them using conventional approach. However, it is difficult to completely restore the image, without the demage of resolution or quality. Removing water droplets from the video is a difficult but very important task at the same time.

For the purpose of solving this problem, this study applies a new methodology to the Generative Adversarial Networks (GAN), to eliminate water droplets distortion it would be better to generate new images which are similar original images, rather than restoring the original images, So that the quality of the marine images restored by GAN will be excellent. However, since the quality of images generated by the existing GAN is not good and can not be created separately from the desired parts, CGAN is applied for the proposed method.

The evaluation is performed by comparing the PSNR. Values of the original images with the restored images using each filter and those generated using GAN.