[1N3-GS-10] Vision, speech: GAN

[1N3-GS-10-01] A New Selective Filtering Method Based on GAN for Removing Unwanted Information from Images

〇Yusuke Machii¹, Yusuke Yamaura¹, Yiou Wang¹ (1. Research & Technology Group, Fuji Xerox Co., Ltd)

Keywords:GAN, Image Processing, Super Resolution, AI, Deep learning

In recent years, image quality has been significantly improved. The currently available digital cameras can produce high resolution images. Such high-resolution images may contain unwanted information for image owner. An automatic filtering method for such information is desired. However, existing methods cause artifact in the processed image. In this paper, we propose a method for solving such problems. We first downsample an image to remove the unwanted information and then recover the image to its original resolution using a super-resolution network. This method can remove unwanted information from the image while maintaining the naturalness of the image. Using the CelebAMask-HQ and Urban100 datasets, we show the filtering performance of the proposal method by quantitative evaluation (PSNR) and qualitative evaluation (Visual check by human).