Classification -based method in optimal image interpolation

Clayton Brian Atkins, Purdue University

Abstract

In this thesis, we introduce two new approaches to optimal image interpolation which are based on the idea that image data falls into different categories or classes, such as edges of different orientation and smoother gradients. Both these methods work by first classifying the image data in a window around the pixel being interpolated, and then using an interpolation filter designed for the selected class. The first method, which we call Resolution Synthesis (RS), performs the classification by computing probabilities of class membership in a Gaussian mixture model. The second method, which we call Tree-based Resolution Synthesis (TRS), uses a regression tree. Both of these methods are based on stochastic models for image data whose parameters must have been estimated beforehand, by training on sample images. We demonstrate that under some assumptions, both of these methods are actually optimal in the sense that they yield minimum mean-squared error (MMSE) estimates of the target-resolution image, given the source image. We also introduce Enhanced Tree-based RS, which consists of TRS interpolation followed by an enhancement stage. During the enhancement stage, we recursively add adjustments to the pixels in the interpolated image. This has the dual effect of reducing interpolation artifacts while imparting additional sharpening. We present results of the above methods for interpolating images which are free of artifacts. In addition, we present results which demonstrate that RS can be trained for high-quality interpolation of images which exhibit certain characteristic artifacts, such as JPEG images and digital camera images. We also present results of a new interpolative image coding method which uses RS along with the well-known JPEG compression scheme. These results demonstrate that for relatively low bit rates, the RS-based compression scheme can improve upon JPEG compression used alone, in terms of subjective image quality (for an approximately fixed bit-rate), and in terms of better rate-distortion tradeoff.

Degree

Ph.D.

Advisors

Allebach, Purdue University.

Subject Area

Electrical engineering

COinS