Design And Implementation Of Image Matching System In Application For Multi-target LED Chips Locating

LED lighting technology is becoming mainstream depending on its excellentfunctionality, safety and comfort. The central part of LED lighting is LED chip.LED die bonder, as a critical process machine in LED production, works on locating,picking and sticking process of LED chips production.For accurate location, detection and classification of semi-finished LED chips,image matching system plays an essential role in the LED die bonder, determiningits performance and quality. This thesis researches the feasibility of optimizationand parallel hardware acceleration of image matching algorithm from two aspects,the matching correlation function and fast serearch strategy, and presents aoptimized image matching algorithm design applied in multi-target locating. Thewhole research includes:Firstly, based on the conditions including multiple targets, sub-pixel accuracyand real-time requirements, the study analyzed several image matching algorithmsbased on respectively frequency-domain and spatial-domain characteristics; andextracted a fast matching method fully exploiting the existing hardware potentialand possessing possibility for further optimization.Furthermore, combined the frequency-domain and spatial-domain algorithm,this thesis presented a matching method based on phase correlation, Fast Fouriertransform (FFT) and polar coordinate transform. A denominator-normalized powerspectrum is taken to achieve multi-target locating, an improved Fourier-Mellinmethod is taken to calculate the angle of rotation, and a two-dimensional surfacefitting method is taken to figure out coordinates in sub-pixel accuracy.Finally, this thesis presented a well-designed functional image matching systemby implementing the design methods and programming techniques of high-speedimage matching algorithm, optimzing core computing parts of the matchingalgorithm of the FFT operation and multi-peak search algorithm, designed forhardware parallel optimization with the Intel SSE2instruction set and the NVIDIACUDA platform. The experimental results show that our developed image-matchingalgorithm system adapts to the rotation interference of targets, and meets thereal-time and accuracy requirements of real application.