PROJECT TITLE :

Learning Multiple Linear Mappings for Efficient Single Image Super-Resolution - 2015

ABSTRACT:

Example learning-based mostly superresolution (SR) algorithms show promise for restoring a high-resolution (HR) image from a single low-resolution (LR) input. The foremost common approaches, however, are either time- or house-intensive, which limits their practical applications in many resource-restricted settings. During this paper, we tend to propose a novel computationally economical single image SR technique that learns multiple linear mappings (MLM) to directly rework LR feature subspaces into HR subspaces. In specific, we tend to 1st partition the large nonlinear feature area of LR images into a cluster of linear subspaces. Multiple LR subdictionaries are then learned, followed by inferring the corresponding HR subdictionaries primarily based on the idea that the LR-HR options share the identical illustration coefficients. We have a tendency to establish MLM from the input LR features to the desired HR outputs so as to achieve quick nevertheless stable SR recovery. Furthermore, so as to suppress displeasing artifacts generated by the MLM-based mostly method, we have a tendency to apply a fast nonlocal suggests that algorithm to construct a straightforward however effective similarity-based regularization term for SR enhancement. Experimental results indicate that our approach is each quantitatively and qualitatively superior to alternative application-oriented SR ways, while maintaining comparatively low time and area complexity.