Research On Key Techniques For Image And Video Compression

With the rapid developmental demand on image, video and multimedia operations of Next Generation Network (NGN), the 3rd Generation Mobile Communication (3G), Beyond 3G and Next Generation Broadband Wireless, enhancement of hardware processing platforms and fall of the cost of memorizers, requirements of the image and video coding technologies which have the high efficiency and speed of coding and can make network environment stable is rising day by day. In recent years, with the advances in CMOS chip techniques, portable mobile terminals are developing rapidly. Because of a lack of special hardware support, they requires fast and memory-saving algorithms of image and video processing. The image and video coding technologies have been studied for several decades and obtained splendent achievements. However, these achievements still cannot satisfy the People's requirement. There are a lot of works still should be done to improve the Performance of image coding. Image and video coding and transmitting technologies are the frontier problem of information science and technology at present, and are paid attention to widely in both theory and application.Supported by the National Nature Science Foundation Project of China 'A study on the fast image orthogonal transform algorithm and parallel processsing of multiprecision and multilayer' and Natural Science Foundation Project of Shandong Province 'Research of Optimization of Realization of International standards on Image', this thesis analyses the application requirement and difficult and key problems existing in image coding technologies and popular international standards of image/video compression and transmission, and conduct systemic and comprehensive research on the key techniques. The main contributions of this thesis are described as follows:(1) Research on the fast computation model of two-dimension forward and inverse Discrete Cosine TransformA new fast two-dimension 8x8 Discrete Cosine Transform (2D 8x8 DCT) algorithm for image coding is presented. DCT is the most popular transform method at present in image compression. The new algorithm reduces the arithmetic operations of DCT by using the specialities of the 2D DCT basic images and computes every DCT coefficient in turn more independently. Hence, the new fast algorithm is suitable for 2D DCT pruning algorithm of pruning away any number of high-frequency components of 2D DCT.An adaptive pruning DCT algorithm for image coding is presented. Based on the above algorithm and a model predicting non-zero DCT quantized coefficients, only the non-zero DCT quantized coefficients will are chosen to compute adaptively. By using several standard images, the experiment results showed that the number of those omitted non-zero coefficients in the new algorithm is very few (only 0.156 per 8x8 block averagely) and they are in the higher frequency place and their values are equal to±1 generally. The comparison of the new algorithm with Feig scaled algorithm shows that the arithmetic operations of DCT is greatly decreased with negligible image-quality degradation.A new 2-D 8x8 DCT algorithm and a predicting model are presented in this paper. We combine Both of them to propose an adaptive pruning DCT algorithm. The adaptive pruning algorithm only computes the non-zero DCT quantized coefficients, hence the arithmetic operations of DCT is greatly decreased with negligible image-quality degradation. The comparison of the new algorithm with Feig scaled algorithm shows that the adaptive pruning algorithm is efficient. Moreover, in the case of the low bit-rate image transmission, the new algorithm is more efficient.A fast algorithm for computing 8x8 two-dimensional Inverse Discrete Cosine Transform (IDCT) is presented. The new algorithm reduces the arithmetic operations by three techniques. The first one is to use the symmetry of the basic images; the second is to combine two steps of computing inverse quantization and IDCT into one, and the last one is to utilize the characteristics of the data of the practical images that the values of the most quantized DCT coefficients of the images are zero, and the values of quite a few of the nonzero quantized DCT coefficients are±1. The theoretic analysis shows that the arithmetic operations are reduced considerably by the combination of the three techniques. By using standard images, the new algorithm is compared with Feig's algorithm that is the most influential one nowadays. The results indicate that the new algorithm decreases about 60 percent of multiplying operations and about 15 percent of addition operations.(2) Research on the multiplierless fast computation model of two-dimension forward and inverse Discrete Cosine TransformThe research is done because of the demand on avoiding multiplying operations in many applications and the points that the cost of memorizers is less and less, and the size of memorizers is smaller and smaller.An 8x8 2D DCT multiplierless fast algorithm is introduced. In the proposed algorithms, the number of addition operations for the transform is also reduced while multiplying operations are eliminated by using Look-Up Table (LUT). By designing skillfully the structure of LUT, we can get a group of correlative product data every time in looking up the table so that the number of looking up the table is reduced greatly. By using the symmetries of basic images and studying the ranges of data in computing the transform, the size of LUT is decreased. In image compression, we can only compute the DCT coefficients which will be encoded and transmited by adopting the proposed algorithm so as to reduce the arithmetic operations more.A new multiplierless fast 2D 8×8 IDCT algorithm based on LUT is presented, in which, the LUT structure is based on the DCT basic images. The LUT size is decreased by using two techniques. One is to use the symmetries of the basic images; another is to elicit the range of each quantized DCT coefficient. If the quantization matrix has the symmetry q(u,v)=q(v,u), the LUT size can be decreased half nearly again. The new algorithm avoids multiplications by using LUT and decreases additions greatly by utilizing the characteristics of actual images and the symmetries of the basic images.(3) Research on relationships among important stages in image compression and building a fast computation modelA novel fast algorithm for color space conversion computation between RGB and YUV is proposed. By combining color space conversion and quantization, the algorithm reduces the multiplication amount greatly in compression and decompression process. This algorithm doesn't change the structure of each processing stage during the image compression, so it has no influence on adoption of other optimization technologies. The data coding/decoding speed is raised efficiently.An optimized algorithm for color space conversion between RGB and YCbCr is proposed. The new algorithm reduces the arithmetic operations of computing color space conversion considerably by combining three stages in an image coder/decoder: color space conversion, quantization and DCT. This algorithm doesn't change the structure of each processing stage, so it is without any influence on adoption of other optimization technology. In each case of image down-sampling, the algorithm raises the image coding/decoding speed efficiently in compression and decompression process.(4) Research on the fast computation model of 2-D 4×4 inverse integer transform algorithm for H.264A new fast two-dimensional 4×4 (2-D 4×4) inverse integer transform algorithm is presented. The algorithm takes advantage of the symmetries of basic images and makes use of the characteristic of the transform data of practical videos. Compared to other fast algorithms, the number of addition and shift operations for the inverse transform can be reduced greatly. Numerical results on several standard video clips indicate that on average, for a 4×4 block, the new algorithm needs 12.7838 addition operations and 1.69536 shift operations, which are much less than the other algorithms need. Moreover, the new 2-D algorithm can be parallelized easily.