Study of Bio-Simulation on Binary-Tree Multiprocessor

(1)对1989年开发的并行语音识别系统进行了改进，增加了DSP（数字信号处理器）进行相关和DP匹配的计算。通过这种改进，并通过去除冗余计算，我们的识别速度提高了20倍以上。由于需要增加滤波器组的数量以获得更高的识别率，因此所需的计算时间将被这种改进所抑制。(2)在二叉树多处理器Coral 68K上实现了基于生物仿真的手印汉字字符识别系统，并对其性能进行了评价。在这个系统中，许多处理器从输入的汉字中提取许多不同的笔画，并通过相互通信，找到具有各自字典中描述的笔画配置的字符。在我们的实验中，超过90%的小于五笔画的手写汉字被识别成功。该方法将适用于图像识别系统。(3)布线问题是将分布在一个平面上的所有端子成双地用互不相交的导线进行连接。当有许多终端时，寻找导线路径的计算时间变得非常大。由于布线是一个强顺序问题，其并行处理一直被认为是非常困难的。因此，我们将生物模拟应用于这一问题，并开发了一种新的并行处理方案，我们称之为竞争处理器并行处理。该算法在Coral 68K上实现，取得了令人满意的性能。

(1) We have improved the parallel speech recognition system developed in 1989 by adding DSP (digital signal processor) to carry out computations for correlation and DP matching. With this improvement and, by removing redundant computations, We obtained more than 20 times increase in recognition speed. As it is necessary to increase the number of filter banks to attain higher recognition rate, the required computation time will be suppressed with this improvement.(2) We have implemented a hand-printed Kanji character recognition system based on biosimulation on our binary-tree multiprocessor Coral 68K and evaluated its performance. In this system a number of processors extract many different strokes from an input Kanji character and, by communicating each other, find the character that has the disposition of the strokes as described in their individual dictionary. In our experiment more than 90% of hand-written Knaji characters composed of less than five strokes were recognized successfully. This method will be applicable to the image recognition system.(3) The wire-routing problem is to connect all terminals in pair which are distributed on a plane with the wires which never intersect each other. When there are many terminals the computation time to find the paths of the wires becomes prohibitively large. As the wirerouting is a strongly sequential problem, the parallel processing of it has been considered very difficult. Therefore we have applied the bio-simulation to this problem and developed a new parallel processing scheme which we call parallel processing with competing processors. This algorithm is implemented on Coral 68K and a satisfactory performance was obtained.