PROJECT TITLE :

Joint Learning of Multiple Regressors for Single Image Super-Resolution

ABSTRACT:

Employing a world regression model for single image super-resolution (SISR) generally fails to provide visually pleasant output. The recently developed native learning methods give a remedy by partitioning the feature house into a range of clusters and learning a simple local model for each cluster. But, in these ways the space partition is conducted separately from native model learning, which results in an abundant range of native models to attain satisfying performance. To address this problem, we propose a combination of specialists (MoE) methodology to jointly learn the feature area partition and native regression models. Our MoE consists of 2 parts: gating network learning and native regressors learning. An expectation-maximization (EM) algorithm is adopted to train MoE on a giant set of LR/HR patch pairs. Experimental results demonstrate that the proposed method will use abundant less native models and time to attain comparable or superior results to state-of-the-art SISR ways, providing a highly practical solution to real applications.