Method for mining common visual pattern between images

Abstract

The invention relates to a method for mining a CVP between images. The method for mining the CVP between the images comprises the specific steps of (1) obtaining a set of possible feature matching pairs between the images, calculating similar values between the possible feature matching pairs, then, building an undirected graph G, solving a maximum subgraph of the undirected graph G to obtain the CVP between the images, and formalizing the solved maximum subgraph to form a local optimum problem; (2) determining the solution mode of the local optimum problem; (3) obtaining an initial solution of the local optimum problem by using a candidate pattern initialization optimizing obtaining method; (4) extending the initial solution by using a candidate pattern extension method; (5) obtaining a final solution by using a final pattern generation method. According to the method, the computation complexity problem, the robustness problem and the precision problem are solved through the candidate pattern initialization optimizing obtaining method, the candidate pattern extension method and the final pattern generation method respectively, mining of the CVP between the images is achieved, and the method for mining the CVP between the images can be widely applied to the field of 2D point set matching, the field of target detection and the like.

Description

technical field [0001] The invention relates to the fields of computer vision and image retrieval, and is a method for mining common visual patterns between images based on graph theory. Background technique [0002] Common Visual Pattern (CVP) is the common part of two images that are consistent in visual content and similar in spatial layout, as shown in the attached figure 1 shown in the connection line. Compared with local features, CVP represents and describes images from a higher level, and is a high-level image retrieval / semantic primitive. CVP mining has very important applications, such as target recognition, 2D / 3D point set matching, local approximate image retrieval, database organization and display, etc. [0003] Common visual pattern mining is generally divided into the following two steps: First, extract the local features of the image, and establish the feature matching pair between the two images. The feature matching pair can use direct matching based on ...

AI Technical Summary

Problems solved by technology

Second, the location, shape, and number of common visual patterns, etc., are not known in advance

Among them, spectral analysis-based methods have the best temporal performance, but they are sensitive to noise and can only mine one-to-one CVP between images.

The method based on quadratic programming generally requires strict graph matching constraints, and can only mine one-to-one CVP between images, so it cannot solve the graph matching problem under unrestricted conditions.

In addition, the objective functions of these algorithms are generally very complex, and the time complexity of solving them is very high.

Although the hypergraph matching method considers the relationship between more matching pairs, it cannot effectively add some constraints in the solution process, and has a higher time complexity than conventional graph matching.

Images