Fast Single Image super-resolution algorithm using feature based regression analysis

Single Image super-resolution (SISR) algorithms focus on increasing the resolution of digital images by establishing a relationship between corresponding low and high resolution images. The relationship is derived using redundant information across the different scales of the image or by using priori information from a training set. Matrix based regression methods have shown superior results when applied to this setting, as they use matrix level image information. In this paper, we extend the concept of a matrix operator to a low level feature super-resolution algorithm that performs edge enhancement and de-noising. Although it is competitively effective in resolving edges and other low level features, it is computationally less efficient when compared to image based operators. We propose a novel patch based implementation, in which the decision to implement the costly feature regression is based on the feature components present in that particular patch. Smooth and gradual patches are resolved by image operators and patches with strong features are resolved by feature operators. The proposed algorithm achieves competitive performance efficiency and is verified by experimental results.