Research On Image Matching Based On Local Feature And Its Application On Bonding Chip Equipment

Machine Vision is a technology that detect and analysis object by machine instead byhuman’s eye. With the development of the computer technology, machine vision is widelyresearched and applied in last two decades. Especially in integral circuit manufacture industry,the machine vision is more widely used. As the key technology of machine vision, imagematch is very important in the process of integral circuit manufacture. In order to picking andplacing the chips fast and correctly, the position must be located by image matching. Since thechips are usually translated or rotated, only the translation and rotation is taken intoconsideration in image matching. In this paper, the researches mainly focus on thefeature-based image matching so that achieves the goal of locating the chips with high speedand accuracy in integral circuit manufacture equipment.This paper introduces a modified generalized Hough transform based on local invariantgeometric feature. The algorithm uses angle between gradient directions of two edge point ofimage as feature, which is invariant to rotation, instead of the feature that used in generalizedHough transform (the gradient angle of one edge point). Like the conventional generalizedHough transform, the algorithm consists two stages: off-line stage and on-line stage. In theoff-line stage, for each edge point, another edge point is found in the local area of it, thereafter,two reference tables are constructed: position reference table (PR-Table) and angle referencetable (AR-Table) in the template image by the angle between the gradient directions of twoedge point. In the on-line stage, the position and rotation angle are voted by the PR-Table andAR-Table constructed in off-line stage. After the voting process, the position and rotationangle are indexes of peaks in accumulate array of position and angle. In order to furtherimprove performance of the algorithm, some optimized strategies are introduced in speed,accuracy and stability.After describing the principle of the algorithm, some experiments are done to test theperformance. The result shows that, the error in x direction is less than0.05pixel, the error iny direction is less than0.15pixel, and the error in rotation is less than0.15degree. Theexperiment result shows that the time remains unchanged with difference angle searchingranges, and the average matching time is15ms when the template image is210x210pixelsand the object image is512x512pixels. Finally, the algorithm is applied in the machine visionsystem of LED bonding chip equipment. Under the guidance of the algorithm, the equipmentruns stability for a long time, the rate of correct matching is99.95%.