Research On Hypergraph Matching Algorithm Based On Tensor

As the base and core of image processing tasks,graph matching plays an important role in the development of computer vision and pattern recognition.Hypergraph matching can effectively solve the problem of graph matching,which is widely used in target identification,tracking,and information retrieval,etc.Nowadays,there are many hypergraph matching algorithms,but the solving of hypergraph matching model is usually NP-hard,so many algorithms use the approximate local optimal solution to replace the global optimal solution of the objective function for this problem.At the same time,hypergraph matching will reduce the matching accuracy by the influence of noise,deformation,outlier points and other factors.Therefore,the main purpose of current hypergraph matching is to seek a method with higher matching accuracy.This thesis focuses on the research of hypergraph matching algorithm based on tensor,two kinds of effective hypergraph matching method are proposed.Firstly,the problem of hypergraph matching can be solved by machine learning,hypergraph matching method based on self-paced learning is proposed.Specifically,this thesis constructs the affinity tensor information between two graphs according to the topological relationship of feature tuples,and learns the specific class model's high-order feature by self-paced learning which trains the model from easy to complex.At the same time,a new self-paced regularizer is used make the model more robust,and the transfer model trained from the learning procedure working as a mediator achieves a better hypergraph matching result between two images.Simulation experiments on different datasets show the effectiveness of the algorithm.In addition,other experiments indicate the propose method has a better ability to resist noise,deformation and outliers.Secondly,according to the cyclic consistency of graph matching,a hypergraph matching method based on low rank tensor recovery is proposed.Specifically,we construct the feature information tensor of a single graph by virtue of the topological relationship between feature points,and then construct the feature information tensor of more images by means of the permutation matrix.Meantime,according to the linear relationship and consistency between image features,the multigraph tensor can be considered to be low-rank.Furthermore,the Alternating Direction Multiplier method(ADMM)is used to recover the low-rank tensor representation which could solve the problem of graph matching.Moreover,the matching accuracy of the algorithm by multiple datasets experiments prove the effectiveness of this hypergraph matching method.Finally,the work of this thesis is summarized.At the same time,the shortcomings of current hypergraph matching methods is discussed and the future development trend is forecasted.