A research into advanced and high-speed multi-media understanding based on a parallel processing model

（1）信号处理与符号处理的集成研究了多媒体处理中的模式信息与符号信息之间的接口机制，提出了自顶向下与自底向上的协同处理框架。（2）多媒体理解的并行处理程序设计在并行多媒体理解系统中，多个处理模块协同工作的多智能体系统的思想是非常重要的。根据这一考虑，我们设计了一个多智能体系统的编程语言，它支持并行面向对象的计算。（3）面向多媒体理解的并行结构我们设计了一个面向多媒体的并行处理器KUMP/D（the Kyushu University Multi-media Processor on Datarol-Ⅱ）的基本结构。在本项目中，我们为实时视频建立了专用的并行I/O接口，实现了实时视频信号和处理器单元之间的同步。（4）基于图像样本的模型自动生成作为图像模式识别模型自动生成的框架，我们开发了一种基于图像样本的二维特征生成一类对象模型的机制。该模型被表示在表示图像区域的分层结构并且具有区域的各种特征的分割树中。

(1)Integration of signal processing and symblic processingWe have investigated an interface mechanism between pattern information and symbolic information, which are quite important in multi-media processing, and proposed a framework of cooperative processing of top-down and bottom-up processing.(2)Programming for parallel processing of multi-media understandingIn a parallel multi-media understanding system, the idea of a multi-agent system, in which various processing modules operate cooperatively, is quite important. According to this consideration, we have designed a programming language for multi-agent systems, which supports parallel object-oriented computation.(3)Parallel architecture for multi-media understandingwe have designed a basic architecture of a multi-media oriented parallel processor, called KUMP/D (the Kyushu University Multi-media Processor on Datarol-II). In this project, we have establishedaspecialized parallel I/O interfacefor a real-time video, which realizessynchronization among real-time videosignals and processor elements.(4)Automatic model generation from pictorial examples,As a framework of automatic model generation for pictorial pattern recognition, we have developed a mechanism, which generates a model of a class of objects based on two dimensional features derived from a set of image examples. The model is represented in a segmentation tree representing a hierarchical structure of image regions and having various features of the regions.