The Application Of The Clustering Algorithm Based On Competitive Learning Mechanisms In Image Matching

In the process of traditional feature-based image matching method, it directly searches theclosest features from the feature sets of the reference image for matching with the query feature.But when the number of extracted feature points become too much, which can sharply increase thenumber of matches between the feature points and will reduce the efficiency of the matching. Infact, the features will be presented cluster structures in high dimensional space, and this structureshould hide the classified information of features. This paper classifies the features and excavatesthe cluster structureoffeaturesthroughcluster algorithm. As a result, the feature space is divided toseveral compact subspaces. During the process of matching the feature points, the algorithm onlyneed to search the matched points in corresponding class, which improves the efficiency of imagematching. However, to get better clustering effect,most ofthe clustering methods need to know thereal number of feature classes in advance. The Rival Penalized Competitive Learning (RPCL)algorithm has the ability of selecting the correct number of clusters automatically, but itsperformance is sensitive to the selection of relevant parameter. And its variant algorithm, RivalPenalization Controlled Competitive Learning, is unreasonable that all the rival units are treated asredundant unitsto be penalized.The author deeply study the problems of competitive learning faced, and proposes adiscriminative penalized competitive learning algorithm (DPCL). In the proposed algorithm, thelearning rate of every unit can adaptively adjust during iteration. And the algorithm uses adistinguishable penalization controlled mechanism to discriminate the redundant units and thecorrect units from the rival units. Giving the correct units slight penalization and giving redundantunits heavier penalization, which make this algorithm get exact number of clusters. Then, thispaper uses DPCL to cluster with the feature sets of the two images before feature matching, andsplits the two feature sets to several subsets respectively. According to the clustering result, all ofthe central units are used to build a KD tree, and every subset data be regarded as nodes toconstruct a KD sub-tree. In the process of matching feature points, the paper uses the BBFalgorithmto searchthe closest features incorresponding subset features. Finally, the author verifies the accuracy of this topic with image matching experiment. Andthe experimental results show that the time of features match has been cut down and effect ofmatch improve significantly after introduce the proposed DPCL clustering algorithm in imagematching.