Matching Algorithms Based On Complementary Feature Information And Spatial Information

Matching algorithm plays an extremely important role in the field of computer vision.How to obtain accurate image matching relationship and accurate point set registration algorithm remains an open problem.In this thesis,the matching algorithm based on the complementary feature information and the spatial information is discussed,while the mathematical theory of constructing information complementary model is explored as well,and it is applied to image feature point matching algorithm and non-rigid point set registration algorithm.On the one hand,the performance of matching algorithm with only local feature information often degrades due to the feature ambiguity,on the other hand,the matching results may be completely wrong with only spatial location information.How to effectively combine the two kinds of information for matching is a challenging problem.To conquer above difficulties,this thesis makes various researches with thoughtful analysis,then the detailed contributions and the primary innovations are listed briefly as below:Firstly,the pipeline of matching algorithm is introduced and described.The scale invariant feature descriptors and shape context descriptors are selected as local features,and the common key point feature descriptor generation algorithms together with the matching algorithm measurement criteria of local feature similarity are introduced.Afterwards,two assignment methods of binary correspondence matrix and fuzzy correspondence matrix are compared and analyzed,and the differences and connections between image matching algorithm and point set registration algorithm are elaborated,which lays the foundation for follow-up research on algorithms.Secondly,an image feature point matching algorithm based on scale invariant feature and spatial location information complementarity is proposed.Above all,a complementary Gauss mixture model is constructed based on local feature information and spatial location information.According to the relevant theories,the energy function to be solved is constructed and the optimal solution is obtained by using the expectation maximization algorithm.The matching probability matrix is estimated through the obtained model parameters,and the matching correspondence is selected by the significance of the probability.Eventually,the algorithm framework is applied to the real image feature point matching algorithm,which can accelerate the convergence speed and ensure the accuracy of matching.Thirdly,a non-rigid point set registration algorithm based on shape context feature and spatial location information complementarity is proposed.Neighborhood structure description using weighted shape context feature similarity is adopted,and incorporated into model based on information complementarity for non-rigid point set registration algorithm.To keep the overall motion direction of the point set consistent,the Gaussian radial basis function is used as the non-rigid deformation constraint,while the motion consistency function is used as the structural constraint.During solving the objective function,the position of the point set is adjusted progressively according to the guidance of feature similarity,which can align the shape of the two point sets point by point.