Image Enhancement Using GAN (A Re-Modeling of SR-GAN for Noise Reduction)

The introduction of neural networks and deep learning models in the field of image processing brought significant progress and achievements. These CNN networks are used in many ways in image processing for classification, segmentation and region of interest detection. They are mainly used in real-time applications like face detections, YOLO, resolution enhancement, etc. The major attention of this paper will be in image enhancement which has wide applications in fields like satellite images, surveillance, etc. The resolution enhancement has been solved in the past using arithmetic operations and has addressed the issue to some extent. The techniques like nearby neighbor and bicubic interpolation (Gao in, Opt. Express 19:26,161–26,173, 2011, [1]) were able to enhance images but were not much reliable in the real applications. With the introduction of GAN (Zhang et al., in [2]) into image processing, the problem of image resolution has been tackled to a certain level. The methods such as EDSR ( Lim et al., in [3]), SR-GAN (Ledig et al., in [4]) and VDSR (Kim et al., [5]) have introduced the CNN for addressing this issue. The SR-GAN model is currently the state of art in the field of image resolution enhancement and delivers a PSNR of 29.40 on Set5 dataset and 25.16 on BSD100 dataset. In this work, we are going to take the method of SR-GAN and address the problem that are left unsolved in the SR-GAN method and improves the PSNR and SSIM score with minor changes in the methodology.