NAKAMURA Tomoo UNIV. OF TSUKUBA; INFORMATION SCIENCES AND ELECTRONICS; ASSOC. PROF., 電子情報工学系, 助教授 (70015923)
EBIHARA Yoshihiko UNIV. OF TSUKUBA; INFORMATION SCIENCES AND ELECTRONICS; ASSOC. PROF., 電子情報工学系, 助教授 (00101155)
NISHIHARA Seiichi UNIV. OF TSUKUBA; INFORMATION SCIENCES AND ELECTRONICS; ASSOC. PROF., 電子情報工学系, 助教授 (50026168)
MINOH Michihiko KYOTO UNIV.; INFORMATION SCIENCE; ASSOC. PROF., 工学部, 助教授 (70166099)
OHTA Yuichi UNIV. OF TSUKUBA; INFORMATION SCIENCES AND ELECTRONICS; ASSOC. PROF., 電子情報工学系, 助教授 (50115804)
The purpose of this research is to get a general understanding of parallel processing for image understanding algorithms. Two types of parallel processing systems are used. One is a system in which several processors are tightly coupled through shared memory. The other is a system in which several processors are loosely coupled through system bus lines or local area networks. The typical image understanding algorithms are selected and implemented on these systems. Several characteristics of these algorithms on the systems are measured and evaluated from several view points. The followings are the results of this research.
1. The framework of parallel processing for the closely connected machine is constructed. Each image understanding algorithm is assigned to a separate processor, and the control information among them is transmitted by the process communication function, while image data is located on the shared memory, so any processor can access them without overhead.
2. We built a lo
osely coupled parallel processing system using transputers. We also built a distributed processing system using a RISC machine and a tightly coupled parallel processing system via local area network.
3. One of the most important problems of robot vision is the correspondence problem of stereo images. We selected an algorithm for this problem and implemented it on both tightly and loosely coupled parallel processing systems, and evaluated the efficiency.
4. Consistency labeling algorithm, which is useful in many applications such as shape matching, and back propagation method in neural networks, are parallelized, and their efficiency is evaluated.
5. We constructed some utilities available in a distributed parallel processing system:(1) a message passing function to Common Lisp environment from others, (2) the general purpose parallel image processing system which is modified from the system originally developed at CMU for a sequential machine, (3) a connecting mechanism between Common Lisp environment and image processing library through the shared memory, (4) an distributed system using the loosely coupled parallel processing system and local area network. Less