Optimal Correspondence Model for Image Matching With Multi-order Features

An optimal correspondence model was proposed for solving image matching problems with multi-order features. A multi-order feature of an image refers to any of its first-, second- and third-order feature, which was defined by a simple feature point, an edge linking two feature points and a triangle connecting three feature points, respectively. The optimal correspondence model was a weighted bipartite graph with multi-order feature as its vertex. With this model the weight could be directly computed and the solution can be easily obtained by the Kuhn-Munkras algorithm. Results show that the model has good robustness for video sequence and graffiti images. Even with obvious rotation, scale, and affine transformation, it can produce a relatively accurate correspondence result, which is usually better than the famous Flann and BruteForce algorithms in OpenCV.