Research On Correspondence Selection Techniques For Pixel-level Based Feature Matching

Pixel-level feature matching is a pivotal step in many computer vision and image processing tasks,which can be widely employed in structure from motion,object tracking,image retrieval,and image stitching.The main steps of pixel-level feature matching include keypoint detection,feature description,feature matching and correspondence selection.However,keypoint detection and feature description often suffer from limited accuracy because of the nuisances such as rotation,translation,viewpoint change,and illumination change,which leads to a large number of undesirable matches in initial feature matches.The main task of pixel-level correspondence selection is rejecting false matches from initial matches and generating reliable pixel-level consistencies between two images.Nevertheless,since there are noises in the real world such as scale,rotation,translation,blur,viewpoint change,and illumination change,and the distribution of false matches is irregular,it is still an open problem how to select reliable correspondences facing different challenges.Consequently,this paper presents three correspondence selection methods that improve the robustness against the nuisances from different perspectives.For the issue of previous correspondence selection approaches being sensitive to specific noise,this paper proposes a fusion method that combines the strength of multiple existed correspondence selection methods and improves the robustness against various nuisances.An experimental benchmark is also developed where the performance of the proposed fusion methods and some state-of-the-art methods is comprehensively evaluated.For multiple consistencies between two images,this paper presents a multi-consistency correspondence selection approach based on a game-theoretic framework,which is not only able to selection single consistency,but also effective to recognize multiple underlying consistencies.For the limited reliability of spatially nearest neighbors in the field of correspondence selection,this paper develops a compatibility-specific mining method to dig out consistent neighbors for correct matches,which is aggregated into a deep learning network in order to further classify initial matches into correct ones and false ones.