Massively Parallel Intelligent Preprocessing

在目前的时序计算机系统中，运动图像的实时处理是非常困难的。因此，我们的目标是实现低功耗的大规模并行预处理网络，基于生物视网膜和大脑的视觉机制。实际上，模拟LSI已经被尝试开发，其检测运动并识别运动目标的简单形状。在生物视网膜和大脑中，目标的边缘信息在早期阶段被提取。基于脊椎动物视网膜光感受器的局部自适应和反馈，实现了边缘检测网络。从具有五个数量级的宽亮度范围的图像中检测边缘。功耗比现有技术下正在开发的视觉芯片低两个数量级。然后，以蝗虫和苍蝇的视网膜和大脑为材料，以边缘信号为输入，构建了运动检测网络。接近的速度 关于我们 发现目标。速度检测所需的时间比现有系统短一个数量级。基于仿形视觉系统，开发了一种简单形状识别网络。其结果是，简单的形状，如圆形，三角形和正方形的识别。该网络采用二进制输出和反馈电路来抑制MOS管失配引起的模拟电路的误操作，可以预测在大规模集成电路中可以实现接近物体运动检测与简单形状识别相结合的网络。曲线运动由局部的直接运动组成。因此，开发了一个网络，用于跟踪目标的特定点的运动。另一种运动检测网络被开发出来，其中背景的运动被抑制。对移动的车辆的避碰具有一定的参考价值。预测当在没有扫描的情况下获得大规模并行输出时，检测时间减少1/（像素数）（1/1000）。少

Real-time processing of moving images is ultimately hard in present time-sequential computer systems. Thus, we aimed to realize massively parallel pre-processing networks with low-power dissipation, based on the vision mechanisms of biological retinas and brains. Practically, analog LSIs have been tried to be developed, which detect the motion and recognize simple shapes of moving targets.In biological retinas and brains, the information on the edge of a target is extracted at an early stage. An edge-detection network was realized based on local adaptation and feedback to photoreceptors in vertebrate retinas. The edge was detected from an image with a wide brightness range of five orders of magnitude. The power dissipation was less by two orders of magnitude than that of vision chips under development with present technology.Then, a motion-detection network was developed based on the retinas and brains of locusts and flies, in which the edge signal was inputted. The speed of an approac … More hing target was detected. The time required, for the speed detection was shorter by one order of magnitude than that of present systems. A network for simple-shape recognition was developed based on the vision system of flogs. As a result, the simple shapes such as a circle, triangle and square were recognized. In these networks, a binary output and feedback circuit were adopted to suppress the wrong operation of analog circuits due to the mismatch of MOS transistors.It is predicted that a network for the motion detection of an approaching objects, combined with the simple-shape recognition, can be realized in a LSI. A curved motion consists of local direct-motion. Thus, a network was developed for a trace of motion of specific points of a target. Another motion-detection network was developed in which the motion of a background was suppressed. It is valuable for the collision avoidance of mobile vehicles. It is predicted that a time for the detection is reduced by 1/(number of pixels) (〜1/1000) when a massively parallel outputs are obtained without scanning. Less