NEW IMAGE INTERPOLATION ALGORITHMS BASED ON DUAL-TREE COMPLEX WAVELET TRANSFORM AND MULTILAYER FEEDFORWARD NEURAL NETWORKS

摘要

In this paper, new image resolution enhancement algorithms based on complex wavelet transform and feedforward neural networks are proposed. The wavelet sub-bands corresponding to high-resolution images are estimated by neural networks using low-resolution counterparts. High-resolution images are then reconstructed employing the inverse transform. We take advantage of dual-tree complex wavelet transform, such as approximate shift invariance, substantial reduced aliasing and directional selectivity, to obtain a richer representation of local structures in interpolated images. These properties make the subband estimation process more effective and lead to more accurate reconstruction of texture and edge regions. We also present a simplified version of the proposed algorithm to reduce computational cost without significant performance reduction. Subjective comparisons and objective quality assessments indicate notable improvement over the conventional bicubic and bilinear interpolation techniques and some typical recently proposed methods.