Attractor Image Coding with Low Blocking Effects

In the present work, a relatively novel method called attractor image coding (commonly known as fractal image coding) for lossy image compression is investigated. Roughly speaking, the original image is partitioned into several disjoint range blocks. For each range block, an a ne transformation is constructed to approximate the range block by another part of the same image. Compression is obtained by compactly storing only the descriptions of these transformations. The decoded image is obtained by iteratively applying these transformations on any initial image. However, the attractor coding, being a block-based algorithm, su ers from the usual blocking artifacts which is highly disturbing to human visual system (HVS). The blockiness is mainly due to the independent processing of each block in encoding. Discontinuities may occur across the block boundaries in the decoded image that are smooth in the original image. The problem is more prominent when the bit rate is reduced. In this thesis, two variants of the existing attractor coding are proposed and examined in details. The main goal is to produce a decoded image which su ers less from blocking e ects and they are summarized as follows: First, a novel attractor coding scheme with adjacent block parameter estimations is proposed that exploits the redundancies of block parameters in smooth regions of an image. A criterion called -minimum edge di erence ( -MED) is proposed to select those blocks which can be estimated well from the adjacent blocks. The goal is to v produce a decoded image in which the blocking artifacts is not easily visible to the human vision systems (HVS). Experiments show that the proposed method can produce a decoded image with most blockiness not easily spotted out and, at the same time, achieve similar compression ability as those of variable range block size. Since the previous proposed method still employs disjoint range blocks in the formulation, the blockiness exists in the decoded images. In this work, a novel attractor coding techniques using the partitioned iterated function systems with lapped range blocks (LPIFS) is proposed. Each range block is formulated to lap with its adjacent blocks through a weighting operator which is diminishing in magnitudes towards its boundaries. Range block preprocessing is proposed to compute the transformation parameters in order to reduce the error. No large system of equations as used in previous work is needed. Moreover, the local domain block matching is proposed. We show that this block matching technique is natural for general images and thus reduces the number of bits for specifying domain block addresses. Experimental results show that the image details are preserved and nearly all undesirable blocking artifacts are eliminated.