CNN-Based Local Motion Estimation Chip for Image Stabilization Processing

Abstract

This paper is to investigate a novel design for local motion vectors (LMVs) of image sequences, which are often used in a digital image stabilization (IS) system. The IS technique removes unwanted shaking phenomenon in image sequences captured by hand-held camcorders. It includes two main parts such as motion estimation and compensation. Most of computation power occurs in the part of motion estimation. In order to reduce this complexity, an idea, which integrates an adaptive-threshold method and cellular neural networks (CNN) architecture, is designed to improve this problem. The design only implements the most important local motion estimation with the array size of 19/spl times/25 pixels. Experimental results with HSPICE simulation and CNNUM are shown that the proposed architecture fast searches the location of possible LVMs and has the capability of real-time operations. The complete design has integrated into the total area of 8.1mm/sup 2/ by using TSMC 0.35/spl mu/m mixed-signal process.