A Comparative Study of PnP and Learning Approaches to Super-Resolution in a Real-World Setting

Abstract

Single-Image Super-Resolution has seen dramatic improvements due to the application of deep learning and commonly achieved results show impressive performance. Nevertheless, the applicability to real-world images is limited and expectations are often disappointed when comparing to the performance achieved on synthetic data. For improving on this aspect, we investigate and compare two extensions of orthogonal popular techniques, namely plug-and-play optimization with learned priors, and a single end-to-end deep neural network trained on a larger variation of realistic synthesized training data, and compare their performance with special emphasis on model violations. We observe that the end-to-end network achieves a higher robustness and flexibility than the optimization based technique. The key to this is a wider variability and higher realism in the training data than is commonly employed in training these networks.