Single image super-resolution method based on linear regression and box-cox transformation

Abstract

Image super-resolution is an essential operation in digital image processing and aims at recovering a high-resolution image from one or more low-resolution input images. The process in which a high-resolution image is obtained involves the creation of new and unknown pixels that must be estimated. Super-resolution (SR) algorithms can be classified as multi-image SR and single-image SR. Multi-image SR methods utilize multiple lowresolution image as input, whereas single-image SR methods utilize only one low-resolution image. Traditional methods such as nearest interpolation, bicubic interpolation and bilinear interpolation are all example of single-image SR methods. These methods, while simple, introduces many artifacts in the high-resolution images. Several methods have emerged in the past decade with the goal of obtaining better, artifacts-free high-resolution images. This thesis proposes a novel single image super-resolution method called BCZ(Box-Cox Zoom), that uses a multiple linear regression model and the Box-Cox transform to interpolate the unknown pixels in the high-resolution image. Our proposed method is compared against the classical interpolation methods using the Ultra Eye image data set. The methods are compared by building confidence intervals using the bootstrap method