Objective: The objective of this investigation is to develop and implement adaptive image compression techniques on massively parallel architectures to meet NASA's needs of archival image sharing and video compression.
Approach: An adaptive block compression algorithm has been proposed. In order to achieve higher compression and fully utilize parallel computing resources, an improved version--bottom-up approach--was implemented and tested. Instead of decomposing from the largest possible blocks, original images are divided into blocks of minimal block size. Then, each block is statistically analyzed, and neighboring blocks, which are visually smooth, are melted into larger blocks. Every block is encoded using its size and mean value only to save storage. Furthermore, since images are examined blockwise, it allows a straightforward parallel implementation. Each group of neighboring blocks can be assigned to a processor and perform computation independently to achieve satisfactory speed-up. The concept of adaptive block compression provides a general guideline to explore the spatial redundancy in images. More complicate and efficient algorithms can be developed based on this framework. Presently, we are incorporating vector quantization (VQ) with this scheme. We expect that VQ or other similar techniques can provide better approximation for each block and, as a result, improve the overall performance.
Accomplishments: The proposed bottom-up approach has been tested on a spectrum of standard images; it showed a better compression ratio than the original algorithm. We have also successfully ported the parallel version onto the IBM SP-2 at the Maui High Performance Computing Center. Experiments have been conducted on up to 16 nodes, maintaining an efficiency of 0.87; speed-up is 13.86. The results show the highly parallelizable property of this algorithm.
Significance: Modern space and earth exploratory technology has enabled rapid acquiring of huge amounts of scientific data, among which most are image and video signals. Thus, an efficient image compression scheme plays an essential role in data transmission and storage. The proposed algorithms provide the ability to control the compressed image quality, which is barely done in the other techniques. Most of the transform-based compression methods can not decide the signal-to-noise ratio before finishing compression. Therefore, several trial-and-error runs are often required to achieve desired accuracy, while our algorithm terminates the bottom-up process dynamically according to user requirements.
Status/Plans: Both serial and parallel programs of the algorithms have been implemented and tested. Currently, we are investigating various techniques to incorporate adaptive block coding to improve the performance of our algorithms.
Point of Contact:
Dr. David Y. Y. Yun
Department of Electrical Engineering
University of Hawaii at Manoa
dyun@wiliki.eng.hawaii.edu
808-956-7627 or 808-956-6349
URL: http://spectra.eng.hawaii.edu/About/SPIC/spic.html