Font Size: a A A

A Study On Methods For Features Matching Based On Graph

Posted on:2015-08-31Degree:MasterType:Thesis
Country:ChinaCandidate:Y T WangFull Text:PDF
GTID:2308330461985048Subject:Computational Mathematics
Abstract/Summary:PDF Full Text Request
Feature matching is a fundamental question involving computer vision. Thus, whenever it comes to two or more images, it will definitely concern with matching corresponding features. In recent years, the feature matching method based on atlas has become a new hotspot for its flexibility, low computational complexity and robustness.In 2005, Leordeanu proposed a theory of corresponding feature matching algorithm based on spectral graph theory, which has drawn much attention from home and abroad.This paper studied the deficiencies and limitations of these methods, and proposed how to improve them. The main research contents and results are as follows:1. Through theoretical analysis and experimental testing on corresponding feature matching algorithm which Leordeanu proposed based on spectral graph,two aspects of limitations were found:(1) Such algorithms only consider the largest eigenvalue of the affinity matrix that corresponds to eigenvector.The underlying assumption is that the correct match is a strong connected cluster. However, if there are much clusters in situations,there are only the largest one that can be retained, resulting in the loss of other clusters.(2) If the initial matching error is relatively of high rate, the function of the method will greatly decrease and the reliability decreases as well.2. For the first limitation of this method,we improved it in two aspects. (1)For the affinity matrix, we should not only consider the eigenvalue that the largest eigenvector corresponds to, but take some larger eigenvalues into account to configure the eigenvector that can reflect the re-matching relations between the feature vectors.And in this way,we can analyze the matching relations.(2)By reusing Leordeanu’s methods for many times,we sequentially determined the largest cluster, the second largest cluster and so on,until all clusters were analyzed. Experimental results show that the two improved methods are superior to the original method in terms of multi-cluster.In addition,this method can also be applied to multi-cluster image retrieval to achieve good results.3. For the second limitation of this method,we proposed a feature matching method based on probabilistic reasoning. First, determine some relatively reliable matching features, and then through analyzing how some other features and the matching features are interdependent, increase the number of matching features gradually by applying baysian formulation.Taking this method,error rate of the initial matching is greatly reduced.And the following step is to achieve the final result by using spectral matching method. At last,experimental results proved the effectiveness of this method.
Keywords/Search Tags:Feature matching, Graph matching, Spectral graph theory, Probability voting
PDF Full Text Request
Related items