Research On Image Matching Based On Local Affine Transformation Consistency

Image matching is a fundamental research in the field of image processing and computer vision,and has various real-world applications.The image matching techniques usually build upon extraction and matching of local features in images.Among them,conventional methods mainly assume that there is a globally uniform geometric transformation between the correct feature correspondences of the input images,and thus can only deal with single object matching with fixed transformations.The recent matching algorithms mainly exploit the graph optimization techniques to locate local feature match sets with similar geometric transformations in the image.Compared with conventional algorithms,they can handle matching multiple objects and object deformation better,but are still limited in matching accuracy,robustness and efficiency in certain cases.Based on the existing researches,this paper further studies image matching mainly by exploring the geometric consistency among local feature matches.To this end,we propose an algorithm based on local geometry consistency evaluation,and a density joint transformational-spatial clustering algorithm for image matching.Our work includes the following aspects:Firstly,a brief introduction to the fundamental knowledge involved in the algorithm is presented.It mainly introduces the image matching process,image feature extraction,clustering methods and the concept of symmetric transfer error,which are expected to lay a foundation for the specific algorithm elaborated in the later stage.Secondly,we propose an algorithm based on local geometry consistency evaluation,which is later utilized to identify correct matches and extended to a heuristic image matching approach.The two keypoint locations of each putative match are separately represented by the keypoints of their neighboring matches using local linear embedding,with each representation indicating the geometric layout nearby each keypoint.As correct matches should have affine-invariant geometric layout,an energy function is designed and optimized to evaluate the differences between the local geometric structures neighboring the keypoint pair of each putative match,which is then employed to exclude false matches with a heuristic strategy.In order to improve the recall rate,we expand the refined match set by finding their neighboring matches,each of which is further evaluated for geometric consistency between its two keypoints based on the matches in the refined match set.The matches with good geometric consistency are further added to the refined match set.We alternate between the geometric consistency evaluation and match compensation until convergence.Experiments on standard dataset show the effectiveness of the proposed geometric consistency evaluation algorithm as well as the presented heuristic matching approach.Finally,an image matching algorithm based on joint transformational-spatial space clustering is proposed.An affine transformation can be estimated to map the local interest regions of each putative match.Observations show that the correct matches mapping the same object between images share similar local transformation,while their feature points on the same image are close.However,the wrong matches do not have the two merits.Therefore,in the joint transformational-spatial space,correct matches are expected to appear in clusters,while the wrong matches scatter all around as noises.In view of this,we propose to use the transformational similarity as well as keypoint closeness to collaboratively prune wrong putative matches.The density of each putative match are evaluated using joint kernel density estimation based on the two distances.To obtain matches with similar geometric transformations and close keypoints,density-based clustering is performed in the joint space.Experimental comparisons on various matching tasks show that the proposed matching method delivers better matching performance than existing algorithms.