A fast and accurate purification method of image feature point pairs using structural consistency constraints

Feature-based matching of images is a key method in pattern recognition, but due to noise, out-of-focus, similar texture, et. al., mismatched feature point pairs are usually generated, and easily result in the incorrect final results. Based on the linear transformation of image scenes, this paper proposes a fast and accurate purification method of matched feature point pairs using the structural consistency of the constructed matching Delaunay triangulations. The structural consistency is expressed and uantified from two aspects: distribution correspondence of the corresponding adjacent vertexes, and the similarity of the corresponding triangular units. By taking the two aspects as restraints, the mismatches are screened out and removed dynamically, which can improve the matching foundation and give better support for the further processing for images and targets. Representative images with different imaging platforms and environments are chosen for experimentation. Results show that compared with the existing classical methods, the proposed method has high accuracy, good robustness and fast computational speed for removing the mismatched feature point pairs.