|
| Research Projects (Dr. Lai-Man Po) |
|
| Last Updated: 19 th August, 2002 |
| Home |
| Top |
Today's research and training on digital video processing, multimedia database, computer vision and virtual reality require supercomputing power, large amounts of main memory, large quantities of disk storage, extremely fast graphics performance and high-bandwidth input/output. It is essential to have a supercomputer system to speed up these large-scale computing experiments and program. A low-cost PC-based parallel supercomputer (SuperAbacus) was built in this project for supporting multimedia signal processing research projects. The SuperAbacus consists of 16 nodes and each node is a 450MHz Pentium-III based PC running Linux operating system. All the nodes are connected by a 100Mbps Ethernet switch to a dual-processor host computer with audio/video interface cards and 72 GB of main storage.
The SuperAbacus could provide more than 2 GOPS (billions of operations/second) computing power and all the necessary I/O interfaces for handling real-time multimedia signal processing such as audio/video compression, multimedia database searching, etc. The SuperAbacus may be the most cost-effective supercomputer in Hong Kong, as it only costs HK$380,000.
Recently, a content-based image retrieval system had been successfully implemented on the SuperAbacus. A medium size image database with 5,000 JPEG images requires less than 13 seconds for image searching and retrieval in the SuperAbacus.
Block-based motion estimation is the heart of the video compression problem. It can consume up to as much as 75% of the total computational power of a video codec. In this project, two new fast block-matching algorithms - four-step search (4SS) and adaptive motion tracking block-matching algorithms are developed. Experimental results show that these two new algorithms perform much better than the well-known three-step search (3SS) in terms of the mean-square error measure while require much lower computational power. They are also more robust as compared with 3SS and possess the regularity and simplicity of hardware-oriented features. Experimental results show that these two new center-biased algorithms perform better than the well-known three-step search (3SS) in terms of the mean-square error measure with much smaller computational requirement. They are also more robust as compared with 3SS and possess the regularity and simplicity of hardware-oriented features.
L. M. Po and W. C. Ma, "A novel four-step search algorithm for fast block motion estimation," IEEE Trans. on Circuits and Systems for Video Technology, vol. 6, no. 3, pp. 313-317, June 1996.
J. B. Xu, L. M. Po and C. K. Cheung, "Adaptive motion tracking block matching algorithms for video coding," accepted for publication in IEEE Trans. on Circuits and Systems for Video Technology, 1998.
The design of optimal codebook with the minimum average distortion for vector quantization has been a challenging puzzler which remains to be solved. The key problem for the codebook design algorithms is how to construct codevectors most efficiently to achieve the minimum average distortion. In this project, a minimax criterion based on Partial Distortion Theorem is introduced for obtaining the uniform partial distortion. By incorporating this minimax criterion into the on-line learning mechanism, a new algorithm called minimax partial distortion competitive learning (MMPDCL) is developed in this project. The extensive experiments have demonstrated that the MMPDCL algorithm consistently produces the best codebook corresponding to the least average distortion compared with other well-known codebook design algorithms such as GLA, FSCL, etc.
A typical digitized mammogram requires approximately 38 megabytes of data, and this is a big concern in the storage requirements of a digital mammogram database. Near-lossless compression is essential for achieving high compression of mammograms while maintaining decoded image quality that will not affect the manual or computerised detection of breast cancer. The objective of this project is to develop new near-lossless coding schemes for digital mammogram compression. These new algorithms are developed based on the lossless Lempel-Ziv algorithms. Lossy and multidimensional generalisation Lempel-Ziv algorithms will be developed in this project. To facilitate the development of computer algorithms for the aid in screening and diagnosis, a small size digital mammographic image database with lossless and lossy compressed data will be established on the Internet during this project. That will assistant the development of computer algorithms to aid in breast cancer screening/diagnosis.
| Top |
| Home | Top |