Bio-Inspired Integrated Vision System Based on the Continuous Valued Number System

预计视觉促进将在汽车安全、手持设备和无线成像中发挥重要作用。同时，对成像系统降低功耗和面积的要求也越来越高。因此，需要新颖的解决方案来实现这一目标。传统的机器视觉系统必须在高速下对视野进行采样。它还必须为每一帧存储大量无用的信息，在兆字节的数据范围内。有用的信息隐藏在每一帧的原始数据之下，而这些信息又必须经过大量处理才能提取出来。对于通常用于运动检测的光流测量，由于光流是由空间和时间现象组成的，因此处理任务更加复杂。这意味着光流测量的视觉系统必须设计为测量时空信息，这需要大量的计算任务。与传统的视觉系统相比，专用视觉芯片是包含图像采集和处理的集成电路。在本提案中，这种类型的视觉芯片正在研究中。该芯片将对图像进行预处理，这意味着它将从每个像素中提取有用的数据。使用混合信号（模拟和数字）电路对图像进行前端图像处理。电路和算法的灵感来自昆虫视觉的生物模型。在昆虫中，眼睛元件具有二维拓扑结构，类似于数码相机中的像素。然而，昆虫的神经回路可以适应变化的光线，并有助于探测移动的物体。昆虫视觉模型可以通过模拟电路实现，从而可以在低对比度环境下有效地工作。最重要的是，这种方法允许电路在一个小封装中实现，可以集成到复杂的系统中，并作为传感器网络和手持设备的预处理单元。

It is expected that vision facilitation will play an important part in automotive safety, handheld devices, and wireless imaging. At the same time demands on imaging systems to reduce power and area are increasing. Consequently, novel solutions are required to achieve this goal. A conventional machine vision system has to sample a visual field at high speed. It also has to store large amounts of useless information, in the range of megabytes of data, for each frame. Useful information is buried under the raw data in each frame which, in turn, has to be heavily processed to extract it. For optical flow measurement, which is generally used for motion detection, the processing task is even more complicated since the optic flow is comprised of both spatial as well as temporal phenomena. This means that the vision system for optical flow measurement has to be designed to measure the spatio-temporal information, which requires a high number of computational tasks. In contrast to conventional vision systems, a dedicated vision chip is an integrated circuit that contains both image acquisition and processing. In this proposal, this type of vision chip is under investigation. The chip is going to perform pre-processing on the image, which means it will extract only useful data from each pixel. The front-end image processing on the image is performed using mixed-signal (analog and digital) circuits. The circuits and algorithms are inspired by biological models of insect vision. In insects, the eye element has a 2-D topology that is similar to pixels in the digital camera. However, the neural circuits in insects adapt to variable light and are useful for detecting moving objects. The insect vision model can be implemented by analog circuits, which in turn can effectively function in low contrast environments. Most importantly, this method allows circuitry to be implemented in a small package, which can be integrated into complex systems and serve as a pre-processing unit for sensor networks and handheld devices.