Adaptive image magnification using edge-directed and statistical methods

Abstract

This thesis provides a comparison of two adaptive image magnification algorithms selected from the literature. Following implementation and experimentation with the algorithms, a number of improvements are proposed. The first selected algorithm takes an edge-directed approach by using an estimation of the edge map of the high-resolution image to guide the interpolation process. It was found that this algorithm suffered from certain inaccuracies in the edge detection stage. The proposed improvements focus on methods for increasing the accuracy of edge detection. The second selected algorithm takes a statistical approach by modelling the high-resolution image as a Gibbs-Markov random field and solving with the maximum a posteriori estimation technique. It was found that this algorithm suffered from blurring caused by the general way in which the clique potentials are applied to every sample. The proposed improvements introduce a set of weights to prevent smoothing across discontinuities. The two selected algorithms are compared to the enhanced versions to demonstrate the merit of the proposed improvements. Results have shown significant improvements in the quality of the magnified test images. In particular, blurring was reduced and edge sharpness was enhanced.