Image Enlargement Method-Based Performance Improvement of Motion Estimation and Compression in Advanced Video Coding

• Main Authors: Chih-YuanHsu, 許智淵
• Other Authors: Shu-Mei Guo
• Format: Others
• Language: en_US
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/5k7837

博士 === 國立成功大學 === 資訊工程學系 === 102 === With the rapid growth of multimedia information, video coding standards utilize several new techniques to improve image quality and compression ratio. However, it increases huge computational complexity. This dissertation studies three topics. First, a fast large-scale image enlargement method with an objective evaluation approach is proposed. Secondly, an efficient block-matching algorithm is proposed for motion estimation from the aspects of theory, analysis and experiment. Finally, we present an improved lossless image compression with efficient error value centralization by sign bits. In the first topic, image enlargement is a critical technique in image processing domain, but there are the shifting effect and the blocky effect that occur in traditional DCT for image enlargement. Therefore, a fast large-scale image enlargement method via the improved discrete cosine transform (DCT) is proposed to improve the computational speed and quality of the large-scale enlargement image. The proposed image enlargement algorithm based on DCT saves computational time by multiplication of the DCT matrix, especially for the large-scale image enlargement. Compared to the traditional DCT approach, the improved approach eliminates the image shifting and blocky effects. In comparisons to other interpolation methods, the proposed DCT enlargement outperforms other interpolation methods in edge details because the proposed DCT enlargement considers the global frequency information of the whole image. With the DCT enlargement, it is easy to implement the arbitrary pixel-size-based zooming of an image by employing the suitable size of transform matrix. Illustrative examples show the proposed approach performs higher image quality and lower computation time than other traditional and existing methods. On the other hand, traditional researches use the specific image reduction methods which significantly influence the comparison of enlargement algorithm. And, specific image reduction methods will lead unfair measure of enlargement performance. Therefore, an objective novel evaluation approach, implemented by the benchmark function-based peak signal-to-noise ratio (PSNR), particularly suitable for evaluating the performance of a large-scale enlargement of a small size image is proposed in this dissertation. The second topic is motion estimation, for reducing the temporal redundancy, the development of a fast and effective ME algorithm has been a challenging task for years. When the resolution of video sequences is not high/clear, the reconstructive performance through motion estimation will be limitative. The proposed image enlargement method can provide more sufficient information for motion estimation. Therefore, we proposed an effective block-matching motion estimation algorithm based on image enlargement preprocessing. The proposed algorithm consists of three effective steps: 1) apply the full search algorithm to construct a polynomial interpolation (PI) model from the group with four skipping frames to determine near-optimal global motion vector for the initial search point, 2) perform an adaptive search block-size decision in the interpolation step to refine the motion vector (MV) and 3) use the half-way stop technique to reduce search points. Our experimental results show that the proposed algorithm achieves a maximum speed-up factor of 366.47 with only 0.52% peak signal-to-noise ratio (PSNR) degradation in comparison with the full search algorithm. In the last two decades, there exist many high-performance prediction-based methods that use different coefficients of causal neighbors in order to exploit the relationship of spatial energy to produce a less error image. Besides, more and more researches focus on the accuracy of predictor; nevertheless, the predictor spends a lot of time on finding the best coefficients of causal neighbors. The objective of our research is to propose an efficient and implementable method to improve compression ratio, without increasing extra computation complexity. Here, we present an improved lossless image compression based on the prediction method with the proposed efficient error value centralization by sign bits. The contribution of this dissertation is to centralize error values in a novel way to improves coding performance. Experimental results show that our proposed method achieves higher compression ratio than the context-based, adaptive, and lossless image codec (CALIC) method for the images with many details or slightly regular texture.