Multi-order Features And Linear Assignment Model For Image Matching

In Computer Vision, the research for image matching mainly includes designingexcellent description methods for image feature and matching strategies. Proposingstrong robustness image feature descriptor is not only an effective strategy to gethigher accuracy of matching, but also an important guarantee of defining thesimilarity between two image features which has more obvious discrimination. Givengood image feature description method, the task of image matching is seeking amodel which can obtain consistency of correspondence between two feature sets. Atpresent, for the definition of matching model, there are ideas based on similaritydistance, matching method based on RANSAC and graph matching aiming atoptimization objective function.In order to obtain accurate image matching, Many feature descriptors andmatching methods were proposed. A Feature descriptor refers to multi-dimensionalvector or matrix. Graph matching is to find correspondences between two node sets oftwo graphs such that they are as similar as possible. It not only has the advantage ofeasy implementation, but also has the exploitations of geometry. In this paper, westudied the description of image feature and linear assignment models, which majorresearch work contains the following aspects:1) Combined with multi-order information in image, using of the structureinformation formed by the vertices in graph, we proposed improved method at theaspect of edge and triangle formed by the feature point which can obtain multi-orderfeatures. Experimental results demonstrate that second-and third-order featuresperform better than first-order feature and make it easier to eliminate ambiguity inimage matching.2) By formulating image matching as graph matching problem, we discussedoptimal correspondence model for image matching with multi-order features. In thismodel, we use Truncated Gaussian Kernel to compute the similarity between twofeatures. Proving the theoretical feasibility by optimization objective function in graphmatching model which is more accuracy than the nearest neighbor method. Besides,we can use the structure information of image in this model. Experimental resultsshow that the optimal correspondence model is feasible and can obtain the globaloptimal matching results.3) Through the analysis of the main factors that affect the optimal matchingmodel, as for the automatic learning of similarity measure between two features, we introduced Bundle Methods for regularized risk minimization in machine learningalgorithm. By training model with image data set to optimize objective function, ourlearning model obtained the parameter which is relevant to compute similarity. Finally,we verified the importance of improving the efficiency and accuracy by usingmachine learning methods to determine similarity parameters, according to theexperiment results conducted in several data sets.