블록킹 현상 감소를 위한 투영 기반 후처리 기법

Abstract

This dissertation proposes a novel hybrid adaptive projection onto convex sets (POCS) postprocessing method in Block-based Discrete Cosine Transform (BDCT) coded images and presents a framework, which is a serial-concatenation of constraint sets based on the POCS algorithm and a simple directional filter. The iterative postprocessing approaches that a decoded image is reconstructed by the iterative projections onto several constraint sets which are formed from not only the transmitted data, but also prior knowledge about an original image based on the theory of POCS is addressed for the first time by Youla and Webb [20], [21]. POCS methods use iterative technique in nature and may involve substantial computation time. Compared to other methods, POCS-based methods have a major advantage in that they are flexible in utilizing known constraints of an unknown image in reconstruction. Generally, from Zakhor’s [17] and Yang’s [15] studies, theses constraint sets can be classified into smoothness constraint set (SCS) and quantization constraint set (QCS). SCS is the constraint set that captures the smoothness property of the desired image; that is, the image before compression should display no discontinuity between blocks. On the other hand, QCS is what prevents the post processed image form diverging from an original image due to the projection onto the SCS. The POCS-based approaches are recursive and adaptive to the original images, however QCS requires a burdened complexity resulted from DCT/IDCT computations in each iterative step and in the meanwhile, due to disappointed performance the algorithm becomes more and more complex. There are some classical POCS algorithms such as Yang’s method [15], [16], and Zakhor’s algorithm [17]. Yang’s algorithm [15], [16] is insufficient to alleviate blocking artifacts and the performance is not good enough, owing to the iterative application of the global low-pass filter and implement only to the boundary pixels, Zakhor’s algorithm [17] excessively blurs image details, and Jiu’s method [18] has high complexity and computation time. To avoid all these disadvantages, we propose a novel hybrid technique that demonstrates an excellent deblocking capability with low complex multi-projection-based adaptive postprocessing deblocking algorithm for DCT Coded Images. Firstly, we split the whole image into several pixel vectors and categorize the pixel vectors by an adaptive constraint set depending on the local statistic properties. Secondly, according to the activity of the pixel vectors, we classify these pixel vectors into three types of constraint sets based on POCS which affects the performance of deblocking progresses decisively. Finally, a simple directional filter is introduced to improve the pixel values in the whole image.