How Down-sampling Affects Supervised-Learning-Based Image Super-Resolutions

Recently, supervised-learning-based single image super-resolution (SISR) methods have been more and more popular, owing to their breakthroughs in SR performance for high resolution (HR) and low resolution (LR) image pairs. However, LR images used for model training and performance evaluation are usually down-sampled from HR images by the same method. Thus, the evaluation results may not be consistent to a large extent when the down-sampling kernel is different. The motivation of this paper is to evaluate the robustness of supervised-learning-based SISR methods against different down-sampling kernels and analyze the impact of down-sampling to the supervised training and SR performance evaluation. We use six kinds of down-sampling methods to construct LR images from the same HR images, and comprehensively evaluate eleven popular supervised-learning-based SISR methods including dictionary learning and deep learning. Experimental results show that the SR performance of supervised-learning-based methods is the best when the down-sampling methods of the training data and the test data are consistent. We also collect a publicly available Standard Resolution Target (SRT) image dataset to provide a quantitative basis for subjective evaluation on real data. Our insights may facilitate further development and better evaluation of learning-based SISR methods.