Image Matching Based On Local Invariant Features

Image matching is an active domain in computer vision researches. Although some successes have been achieved over decades, image matching still remains a challenge because of the complexity of imaging conditions. This dissertation mainly focused on how to match a large corpus of multi-view images using local invariant features and their spatial configurations, and the proposed method was applied to image registration and measurment of image similarity.In this dissertation, firstly, the SIFT and SURF feature extraction algorithms based on scale space theory were studied, and were implemented to extract SIFT and SURF feature points successfully. In the stage of local extremum detection in SIFT feature point detection, the points were divided into two types: local maximum points and local minimum points. In this way, SIFT feature points could be used to accelerate image matching based on the distance ratio of the closest neighbor to the second-closest neighbor.As there were some mis-matches after the early stage of image matching based on the distance ratios, two different methods to exclude these mis-matches were explored. Both of them, the two-phase iterative Procrustes matching and the classical RANSAC algorithm, could eliminate some of the mis-matchings effectively. However, the former is very time consuming as the number of initial matchings grows; and the result of the latter is of strong randomicity and uncertainty, and it is also very time consuming when the percentage of initial mis-matches is high. In order to solve the problem, an improved RANSAC algorithm using a precise kernel was proposed, and got the correct matches efficiently. The proposed method was used in registration of images from a common digital camera, as well as in registration for multispectral images of different wavelengthes. Both SIFT and SURF features can be used in the proposed method, and it's practicability were validated.Finally, attributed graph representation models were constructed for SIFT features and SURF features respectively, and the similarity measure between two graphs were defined. With the accelerated image matching method based on the distance ratios, the selection of salient local invariant features could be much faster. Images could be represented by attributed graphs of these salient features. A graph matching algorithm that utilizing both the global and the local feature information comprehensively was proposed, but matching results were not the same as the matching order of the graphs changed. So a two-way matching algorithm for attributed graphs was proposed using our improved RANSAC algorithm. A similarity measure between graphs was defined according to the matching results, which could be used for image modeling and object recognition.