RESEARCH OF NATURAL SCENE IMAGE RECOGNITION SYSTEMS HAVING IMAGE SEGMENTATION FUNCTIONS USING NONLINEAR DYNAMICS

点击翻译按钮获取中文摘要

The final target of this research is to construct an intelligent VLSI system that recognizes and understands real-world scene images with real-time. This system sequentially extracts target image regions from the input image in a time scale of microseconds and performs feature extraction using fast wavelet transformation. In this research, processing models suitable for VLSI systems have been proposed, and prototype LSI chips have been designed, fabricated, and evaluated. We have used resistive-fuse networks for rough region segmentation, nonlinear oscillator networks for region extraction, and nonlinear cellular networks for feature extraction.1. Processing models and algorithms suitable for VLSI implementationWe have proposed an oscillator network model that can segment gray-level images. This model requires only 5 bit calculation precision. We have also proposed a double-thresholding algorithm that can climinate incomplete segmentation timing. We have proposed a model for segmentation and extraction of target regions from natural scene images with arbitrary resolution by combining resistive-fuse and oscillator networks. We have achieved flexible human face recognition from natural scene images by combining these network models, Gabor-wavelet transformation, and a flexible graph matching method based on the dynamic-link architecture.2. Design and evaluation of prototype VLSI chips for image processing(1) By using newly developed pulse modulation circuits for arbitrary nonlinear transformation, we have designed oscillator network chips : (a) die size : 4.5 mm sq., 11x11 pixels, 0.6 micron CMOS technology, (b) die size : 9 mm sq., 50 x 50 pixels, 0.35 micron CMOS technology. The measurement results have confirmed oscillator operation with 100 micro-second.(2) We have designed a 1-dimensional pixel-parallel processing network VLSI chip that can perform resistive-fuse networks and cellular networks for Gabor transformation.