Implementation of Analog Integrated Circuits for Motion-Detection Network Based on Retina and Brain

在人类和动物的大脑中，目标的运动是从视网膜中成像的二维图像中检测出来的。即使是低级动物也能进行的运动检测有望广泛应用于车辆防撞系统、机器人眼睛等。本研究试图建立基于生物视网膜和大脑，特别是低级动物大脑功能的运动检测模拟集成电路的基本技术。首先，设计了一种简单的局部自适应基本电路。通过电路仿真和测试芯片的测试结果证实，在亮度变化较大的情况下，通过改变空间分辨率，目标边缘的信号强度几乎保持在给定的范围内。为了提高由光电二极管构成的输入级的灵敏度，设置了两种类型的放大器。一个是用于非常小的光电流的放大器， 关于我们 用于常规光电流。结果，动态范围增加了七个数量级。其次，基于昆虫的脑功能，设计了一个二维运动检测的模拟网络。通过电路仿真证实，通过检测边缘长度的增长率，可以获得接近物体的速度和方向。在与运动检测电路连接的边缘检测电路中形成二进制输出，防止了模拟计算误差的传播。第三，我们利用树突的时空信息，其中信号的延迟时间取决于突触连接的位置，空间信息由突触连接的空间分布给出。用简单的基本电路构成基本网络。其结果是，它是澄清，通过电路仿真，可以获得的时空信息。边缘检测和运动检测都可以在单个网络中完成。这个基本网络在神经形态视觉系统中是广泛的。少

In human and animal brains, the motion of a target is detected from a two-dimensional picture imaged in their retinas. The motion detection that is given to even low-level animals is expected to be widely applied to systems for the collision avoidance of vehicles, the eye of robots and others. In this study, we tried to establish the fundamental technologies of analog Integrated circuits for the motion detection based on the functions of biological retinas and brains, particularly the brain of low-level animals.Firstly, a simple fundamental circuit for local adaptation was developed. It was confirmed by circuit simulation as well as measured results of test chips that the signal intensity of the edge of targets was almost kept in a given range by varying spatial resolution against widely varied brightness. In order to increase the sensitivity at the input stage constructed with photodiodes, two types of amplifiers were set. One is an amplifier for very small photocurrent and the other … More is for conventional photocurrent. As a result, the dynamic range was increased by seven orders of magnitudes. Secondly, an analog network for two-dimensional motion detection was developed based on the brain function of insects. It was confirmed by circuit simulation that the speed and direction of an approaching object could be obtained by detecting the increasing rate of edge lengths. Forming a binary output in the edge detection circuit, which was connected with the motion detection circuit, prevented the propagation of the error of analog computation. Thirdly, we utilized the spatio-temporal information of dendrites in which the delay time of the signal depends on the location of synaptic connections and spatial information is given by the spatial distribution of the synaptic connections. A fundamental network was constructed with simple fundamental circuits. As a result, it was clarified by circuit simulation that the spatio-temporal information can be obtained. Both of the edge detection and motion detection can be done in a single network. This fundamental network is extensive in neuromorphic vision systems. Less