| The digital image, with the rapid development of network-age and the advantage of vivid, intuitive, intelligible, rich of content, has become an important medium in this era of information transmission. But due to these advantages, the more image data swarms into the network, the more images appear with similar visual content, and the larger information redundancy. The image-based technology of objects matching is committed to identify the meaningful foreground among two or multiple figures automatically, it can reduce the redundancy of information exponentially. Therefore, the image-based technology of multi-objects matching will be the one of the hottest research topics in the field of computer visual and pattern recognition in recent years.Since there are considerable visual objects variations and complex spatial layouts among the two or multiple related figures, which makes multi-objects matching a challenging task. Many methods have been proposed to identify visual objects in recent years, they can be categorized into two groups:bottom-up and top-down. Bottom-up methods mainly investigate local spatial cues of visual objects while lack global information, in contrast, it’s not trivial to learn a robust model parameter and posterior probability inference for top-down methods. According to the analysis of them, we propose a method without complex mode parameter learning and inference, which integrates global and local spatial information.In this paper, we propose an approach of multiple objects matching via nonlinear mean shift by viewing this problem from a novel perspective. Specifically, first of all, we extract local affine region from image data via local feature detector, and then generate feature descriptions with SIFT descriptor, and further establish initial set of correspondences. Followed by estimating the geometric transformation between each pair in the set, all the transformations are stacked together to form a similarity geometric transformation space. The similar geometric configuration between common objects makes visual regions to be similar to each other. It reveals a dense distribution in the underlying transformation space. Because of the differences among visual objects, it will form some independent clusters. Hence, matching multi-visual objects can be achieved via mean shift clustering procedure using mode finding.Unfortunately, the original mean shift algorithm is not directly applicable to the constructed transformation space because of its non-Euclidean nature. Based on the analysis of the properties of the space and of the existing nonlinear mean shift algorithm, we propose a new distance measure between two geometric transformation matrixes in space, as well as a new way to estimate the mean shift vector in the "shift" procedure, which provides better clustering results and convergence performance during the mean shift procedure.Extensive experiments on visual regions matching, multiple objects matching in single or multiple figures as well as near-duplicate image retrieval verify the robustness of the proposed approach. |
