Low-Resolution Image Upsampling Method Using Super Resolution Based Adaptive Pixel Shuffle

Super resolution (SR) methods enable to generate high-resolution images from low-resolution images. Since the advent of deep learning based SR methods, those studies have been actively investigated in a variety of image processing like image restoration, noise reduction, image resizing, and so on. Although most SR methods generally utilize the pixel shuffle as one of the upsampling methods, the arrangement for pixel shuffle always follows a fixed pattern according to the channel depth of final feature maps. In this paper, we propose adaptive pixel shuffle method to enhance the accuracy of SR based interpolation process. The proposed method can achieve improved PSNR gains as high as 0.62 dB on average compared to the existing upsampling method on the Set5 and Set14 test dataset.