| Research Abstract |
Despite the phenomenal advancement in the digital computer technology, human-like intelligent information processing is not yet possible. In this project, we have aimed at building a human-like flexible image recognition system using the state-of-the-art silicon technology based upon the "Psychological VLSI Brain Model" proposed by the head investigator. Taking the medical X-ray picture diagnosis as a test vehicle, we have succeeded in developing an intelligent image recognition algorithm, demonstrating the diagnostic results approved by experts having more than 10 years of experience in a university hospital. Furthermore, the real-time response capability of such systems has been shown by developing VLSI chips dedicated to associative image processing. As a result, we have established a solid foundation on which we can build a real-time-response low-power human-like-brain-computing VLSI system in the future research project.
In perceiving and understanding something, recalling the past
… More
experience most relevant to the current event is occurring in our brains as the most basic processing. Being inspired by such a psychological brain model, we have developed the maximum-likelihood-search VLSI engine, which we call associative processors. Various types of high-performance associative processor chips have been developed using both digital and analog CMOS technologies to meet varying needs. Among them, we have developed a ferroelectric associative memory for the first time as a candidate for use in mobile applications.
The most important achievement in the project is the invention of a new image representation algorithm called PPED (projected Principal-Edge Distribution), which has enabled us to carry out robust image recognition using the associative processors. The algorithm bases on the extraction of most primitive features in the image (which we called "piclet") by a dedicated VLSI chip to form feature vectors. The PPED representation very well preserves the human perception of similarity among images in the vector space while achieving the substantial dimensionality reduction in the image data, thus providing the most favorable feature for hardware recognition systems using associative processors. Less
|
