Designs of Discrete Cosine Transform for Advanced Video Coding

• Main Authors: Chao-Hsuing Tseng, 曾昭雄
• Other Authors: Jar-Ferr Yang
• Format: Others
• Language: en_US
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/60264778611132717506

博士 === 國立成功大學 === 電機工程學系碩博士班 === 94 === 　　In the dissertation, we proposed several fast discrete cosine transform algorithms and integer transforms to reduce the computational complexity and achieve better energy compaction of video coders. First, the fast two-dimensional discrete cosine transform (DCT) and inverse discrete cosine transform (IDCT) algorithms were proposed to reduce computational complexity with regular and modular architecture. Then, a systematic design procedure of integer discrete cosine transforms (integer DCTs) and integer orthogonal discrete cosine transforms (IODCTs) was proposed to achieve better energy compaction and improve video coder performance. Finally, an enhanced rate-distortion cost function was proposed to improve the coding performance for H.264/AVC intra mode decision. The detailed discussions are addressed in the following: 　　The proposed fast DCT and IDCT algorithms by using the direct computation approach are based on regular quad-matrix process. Since the algorithms through decomposition and reconstruction procedures can be repeatedly performed, we can easily extend them for the higher-dimension DCT and IDCT computations. With regularized procedures, all the heavy computations can be realized by the same computational kernel, which demands three multiplications and eight additions for each kernel. With high regular architecture and low computational complexity, the proposed algorithms after feasibility design show their advantages in both software and hardware implementation. 　　The integer transform without drifting problems has been widely investigated. Among these researches, the integer transforms in various versions of H.264/AVC are the most attractive. We proposed a systematic design procedure of integer discrete cosine transforms (integer DCTs) and integer orthogonal discrete cosine transforms (IODCTs). Based on the proposed methods, we can design optimal integer transforms with better compaction ability and less computational complexity. With recursive design method, we can get many IODCTs and their reduced computations. The IODCTs depend on selections of normalization factors and cosine kernel integers. We use the compaction coding gain as the criterion to verify the performance of energy compaction to select a proper discrete transform. We found the famous integer transforms which achieve good approximations of the original DCT suggested in H.264/AVC coder all belong to IODCTs. Simulations show that the proposed IODCTs achieve better energy compaction and coding performances than the original DCT and integer transforms in H.264 coder. With advantages of computational efficiency and energy compaction, we believe that the proposed IODCTs could be efficiently and effectively used in advanced video coding systems. 　　In H.264 advanced video coding (AVC) standard, the intra prediction plays an important role in compression of intraframes by referring surrounding coded blocks. It is obvious that either the SAD or SATD criterion suggested in the reference software will cause the worse coding performance compare to RD-optimized criterion. We first propose an enhanced cost function for intra 4x4 mode decision in H.264/AVC and then develop fast computation algorithms of the SATD and the SAITD to reduce the computation by using the property of linear transform and fixed spatial relation of predicted pixels in each intra mode. Simulation results show that when we adopt the enhanced cost function to select the best mode, the coding performance is better than the SAD (or SATD) criterion and is very similar to the RD optimized criterion in low bit rate. Moreover, with the developing fast algorithm of the SATD, we can reduce about 54% computation of the original SATD algorithm for intra 4x4 mode decision. And we can further reduce about 30% the computation of the original SAITD algorithm when computing the enhanced cost function.