A Video Optical Interface Architecture for Neural Network Image Processing
Abstract
In the near future the prospect of using CNN processors in conjunction with CCD video imagers is good. CCD's may soon have on-chip neural networks capable of performing quick and valuable processing of the image before it is scanned. In spite of that, it is unlikely that the first dual imager/processor chips will use charge coupling. Instead, the first video imaging/processing chip will probably be a pin for pin neural network replacement for an existing CCD product. Rather than integrate the charge storage and transfer methods of CCD's, the prototype chip will likely retain the isolated photoconducting pixels associated with current neural network architectures. This scenario will allow the network to be thoroughly tested as a video imager/processor. Afterwards, a scheme will be designed to integrate neural network architecture with the channel MOS structure associated with CCD's. Thus CCD charge transfer methods will replace the discrete pixels and allow the chip to be fully compatible for Raster scanning.
The incorporation of analog processors into video technologies will require an interface which is compatible with existing CCD chip designs. This paper describes an initial video interface architecture which has been designed for current neural network processing systems. The on-chip interface translates a processed image from a network array to a serial digital data stream. The data is then carried off chip to a computer screen for display. This paper details the image collecting and processing features and notes the utility of this video interface as an important step towards a dual imaging/processing IC.
The next two sections explain relevant image processing information to the reader. Section II, describes optical imaging by looking at both the charge coupled device and CMOS photodetecting pixels. Section III describes image processing with "smart-pixel" neural networks. In section IV the video interface is explained including the CMOS chip design and the hardware and software necessary for test and operation. Conclusions and acknowledgments follow.
Description
Program year: 1994/1995Digitized from print original stored in HDR
Subject
cellular neural networkcharge-coupled device
dual imager/processor chips
video interface architecture
image processing
Citation
Forthman, Christopher L. (1995). A Video Optical Interface Architecture for Neural Network Image Processing. University Undergraduate Fellow. Available electronically from https : / /hdl .handle .net /1969 .1 /CAPSTONE -RooneyT _1990.