Multi-scale Information Distillation Network for Image Super Resolution in NSCT Domain

Deep learning based methods have dominated super-resolution (SR) field due to their remarkable performance in terms of effectiveness and efficiency. In this paper, we propose a new multi-scale information distillation network (MSID-N) in the non-subsampled contourlet transform (NSCT) domain for single image super resolution (SISR). MSID-N mainly consists of a series of stacked multi-scale information distillation (MSID) blocks to fully exploit features from images and effectively restore the low resolution (LR) images to high-resolution (HR) images. In addition, most previous methods predict the HR images in the spatial domain, producing over-smoothed outputs while losing texture details. Thus, we integrate NSCT and demonstrate the superiority of NSCT over wavelet transform (WT), and formulate the SISR problem as the prediction of NSCT coefficients, which is able to further make MSID-N preserve richer structure details than that in spatial domain. The experimental results on three standard image datasets show that our proposed method is capable of obtaining higher PSNR/SSIM values and preserving complex edges and curves better than other state-of-the-art methods.