Image Feature Point Matching Algorithm Research

Feature points matching plays an basic and important role in computer vision and pattern recognition. It can be widely used in many areas such as image registrtion,object recognition,motion target detection and tracking,handwritten characters recognition and so on.At the same time ,it is also an difficult problem in computer vision.Because the tow point sets not only have different degree of noise but also exist in outliers and non-rigid transformation,which make the issue of points matching somewhat difficult and complicated.Feature points matching intents to determine correspondence and transformation between two sets of points in space.It is the simplest form of features that point features represented by the point coordinate.In a great number of pratical applications,there is non-rigid transformation for points matching. So in this thesis, two applied feature points matching algorithms are presented in allusion to non-rigid feature points matching:(1) A based on finding invariable approach for feature points matching is proposed which improves Chang's points matching algorithms.The new approach uses finding matching pairs support method of original algorithms,and then according as distance invariable of two points between one set and another set from top to bottom gained maximum matching pairs set and affine transformatin parameters.Theoretical analysis and experimental results show that the improved algorithm has rigorous consequence and better accuracy and stability which is a perfect feature points matching algorithm.(2) The second means of algorithm improve based on deterministic annealing point matching approach. The improved algorithm present the new form of target function which fits for calculation of deterministic annealing method ,and adopt the comparability measuring value as restriction of matching matrix in order to accelerate convergence of matching matrix and reduce quantity of calculate, when anneal ratio is relative higher,improved algorithm can also receive perfect results.Experimental outcomes validate that improved method's reasults are better than original algorithms under the same anneal ratio and can attain perfect matching reasults all the same when anneal ration is reduced befittingly.