Image upsizing with adaptive Wiener filtering method using self-prediction

In this paper, a straightforward and effective method for image upsizing is presented; upsizing is considered to be an essential operation in image processing. The Wiener filtering method is a well-known optimal framework for signal prediction, which provides information of original and degraded signals. In image upsizing, the Wiener filtering framework is not valid because of the missing information about the upsized target image. The proposed method employs a self-prediction scheme where a down-sampled image of the original image is used for the Wiener filter prediction and the obtained filter coefficient is actually used in the upsizing of the original image. The experimental validation shows improvement in both visual and objective quality.

I. Introduction

Image upsizing is a very essential process of image processing, and various applications adopt this technique to meet their own purpose, such as zooming in on a digital image or resealing a video sequence into the display resolution to fit the screen boundary. Thus far, many image upsizing methods have been proposed in the literature [1]–[7]. The most common of these methods is the cubic convolution method [1], where a linear filter kernel was modeled as a piecewise cubic polynomial to interpolate the sampling surface. Many results show significant improvements as compared to the simple bilinear method, while reducing blurred details and ringing artifacts around edges. However, a new method proposed in [2] preserves edge directionality when local characteristics of the image domain are modeled by using a covariance map. The back-projection [3] method minimizes the reconstruction error with an iterative procedure. Another approach for image upsizing is the use of super-resolution methods [4]–[6], where multiple images are used for registration and sampling grid construction. However, all methods noted here suffer from certain inherent problems such as complexity, excessive blurring, and enlarged noise.