A Progressive Training Framework for Image Super-Resolution

A deep network that is stacked has good performance in single image super-resolution (SISR), but it has problems training difficulty and is time-consuming. In this paper, we propose a progressive training framework for SISR, called Progressive Multi-Granularity Training on Super Resolution (SRPMG). Progressive training can help stabilize the training quality and improve the model performance. The SRPMG framework cuts the training images into various sizes and lets different parts of the CNN-based SISR model train on these images so that the model can focus more on different granularity features in each training stage, improving the Recovery quality of edge information. Our method is a plug-and-play method that can be generalized to most CNN-based super-resolution models. Experiments show that our method has a certain improvement among 7 different models, making their reconstructed images have better performance on edge information. At the same time, SRPMG allows these models to improve their results on the 4 benchmark datasets by a maximum of 0.44 DB on PSNR.