Download files
Abstract
As the amount of information is increased in modern life, the need for more efficient methods to represent this information grows as well. This information can be in a variety of forms, such as speech, still image, text, and video, etc. The aim of data compression is to provide efficient methods to represent information. Data compression is very important in the storage and transmission of digital image and video. According to the application, data compression can be classified into two areas: lossy compression and lossless compression. Lossy compression can provide highly compressed ratios than lossless compression, but at the cost of some loss information. Lossless compression is applied in some particular applications for which the compressed data must be reconstructed to be identical to the original information. This thesis presents new approaches to lossless image compression, applied in binary image compression (i.e. binary shape coding) for object-based video compression, and grey-scale image compression. In a binary shape video sequence, each the pixel has value '1' within an object and value '0' for the background, where each pixel is either completely inside the object or completely outside it, i.e. there is no blending of pixels at object boundaries. The main purpose of binary shape coding is to use fewer bits to describe the object boundary information. There are many types of coding methods in the binary shape coding. Most of them exploit the high correlation between adjacent pixels to provide high coding efficiency. In this thesis, we propose a new approach to binary shape coding based on quad-tree block-based CAE coding. The original image is decomposed into different new images with different resolution. Coding is then performed on 3-symbol quad-tree blocks instead of the original binary pixels. The quad-tree structure can remove the spatial redundancy efficiently. The low resolution image costs fewer bits to locate the boundary block information. High resolution image gives more details around the boundary block. This technique can be used in either an Intra mode (i.e. no information from other frames is used) or an Inter mode (in which information from the previous frame is used to further improve the coding efficiency). The efficiency of quad-tree structure is highly dependent on the complexity of shape and motion of the object. An optimal pruning quad-tree prunes the inefficient branches in the quad-tree structure and has most efficient expression for all binary shapes. A fast and simplified algorithm of optimal pruning quad-tree coding is also proposed in this thesis. In comparison of MPEG-4 CAE, the proposed approach gives significant improvement in both Intra and Inter coding. Lossless image compression typically includes two parts: transformation and coding. For reasons of computational efficiency, the transformation must be linear and integer. In natural image, the content in a small region has high correlation, which means lots of redundancy can be removed. Predictive models tailored to exploit redundancy in a particular direction in each block are applied to remove the spatial redundancy. Pixel values in a small size block can be re-arranged in the new range rather than [0, 255]. Variable length coding assigns automatically the short code words to high-probability values. According to the frequency distribution in the local block, small blocks with different patterns are designed to achieve high coding efficiency. Adaptive context-based arithmetic coding can be added for a high-performance compression algorithm, which is applied to the previous low complexity coding results. The results give minor improvement in comparison of other compressions.