Context-based adaptive image resolution upconversion

Abstract

We propose a practical context-based adaptive image resolution upconversion algorithm. The basic idea is to use a low-resolution (LR) image patch as a context in which the missing high-resolution (HR) pixels are estimated. The context is quantized into classes and for each class an adaptive linear filter is designed using a training set. The training set incorporates the prior knowledge of the point spread function, edges, textures, smooth shades, etc. into the upconversion filter design. For low complexity, two 1-D context-based adaptive interpolators are used to generate the estimates of the missing pixels in two perpendicular directions. The two directional estimates are fused by linear minimum mean-squares weighting to obtain a more robust estimate. Upon the recovery of the missing HR pixels, an efficient spatial econvolution is proposed to deblur the observed LR image. Also, an iterative upconversion step is performed to further improve the upconverted image. Experimental results show that the proposed context-based adaptive resolution upconverter performs better than the existing methods in both peak SNR and visual quality.