Ensemble CNN in Transform Domains for Image Super-resolution from Small Data Sets

Abstract

In this paper, we develop a method to recover high-resolution images from small training data sets by combining ideas inspired by compressive sensing, super-resolution neural networks, and ensemble learning. The key idea is that the frequency domain, and other transformed domains, where images can be presented sparsely can be used to ensemble a robust model and improve the performance of super-resolution convolutional neural networks. Fourier and wavelet representations are popular candidates since they provide sparse image representations but, the selection of an optimal transform domain is challenging since each type of transformation may be superior for different sub-image patches in different application domains. Herein, we propose an ensemble method for robustly selecting the appropriate sparsifying transform from limited available data. Our proposed ensemble CNN decides the best choice of transform domains for each sub-image patch and combines the patches to provide a general optimal solution for the entire image.