Super-resolution reconstruction using deep learning: should we go deeper?

Super-resolution reconstruction (SRR) is aimed at increasing image spatial resolution from multiple images presenting the same scene or from a single image based on the learned relation between low and high resolution. Emergence of deep learning allowed for improving single-image SRR significantly in the last few years, and a variety of deep convolutional neural networks of different depth and complexity were proposed for this purpose. However, although there are usually some comparisons reported in the papers introducing new deep models for SRR, such experimental studies are somehow limited. First, the networks are often trained using different training data, and/or prepared in a different way. Second, the validation is performed for artificially-degraded images, which does not correspond to the real-world conditions. In this paper, we report the results of our extensive experimental study to compare several state-of-the-art SRR techniques which exploit deep neural networks. We train all the networks using the same training setup and validate them using several datasets of different nature, including real-life scenarios. This allows us to draw interesting conclusions that may be helpful for selecting the most appropriate deep architecture for a given SRR scenario, as well as for creating new SRR solutions.