Research On Feature Correspondences Of Fisheye Image Based On Geometric Constraint

The process of image feature correspondences is actually searching for one-to-one feature mapping of two images which is from the same scene of space. The image feature correspondence is a key step in the other computer vision problems, for example, three-dimensional reconstruction, target detection, scene analysis, objects recognition and so on. At present, feature correspondences of image has a widely range of applications in practice, it is widely used in such field as medicine, aerospace, biotechnology, information processing and so on. It has become an integral part of digital image processing. In this paper, the main work is focused on feature correspondence based on geometric constraint. Our principle tasks are as follow:1. Fisheye image as an example, we present an efficient method for feature correspondences of fisheye image and plane detection based on these feature correspondences, which exploits hierarchical clustering. And during the process of clustering, we introduce an adaptive partial linkage model. The method takes into account both photometric similarity of correspondences and pairwise geometric constraint of local feature. It calculates pairwise geometric dissimilarity through the homography of feature correspondences, which to measure the geometric consistency of pairwise feature correspondences. Then get the reliable feature correspondences of fisheye and plane detection based on these reliable matches. In our experiment, we compare adaptive partial linkage model to single-link model and the experimental results show that the adaptive partial linkage model is more efficient than other linkage model.2. Aim at there are a lot of false matches in the initial feature correspondences of fisheye image, we present an efficient method based on geometric constraint for removing the wrong matches, which exploits spectral technique clustering. The method takes into account both photometric similarity of correspondences and pairwise geometric constraint of local feature. This method firstly supposes that the image meet affine transformation locally. And then estimate geometric constraint of pairwise feature correspondences, which is used to construct the adjacency matrix of spectral clustering. Finally, remove wrong matches from the initial feature correspondences through the spectral clustering method. In our experiment, we use this method in different scenes, such as common scene of indoor, three-dimensional scene of outdoor, the scene exciting larger deformation, the scene containing repetitive texture. The experimental results show that the method based on geometric constraint, which exploits spectral clustering, can apply to different scenes better and can remove wrong matches from initial feature correspondences efficiently, then get reliable feature correspondences of fisheye image.