Research On Multi-Scale Feature Based Video Image Mosaic

Video image mosaic is a very important issue of image processing in computer vision,which can construct high resolution image with large view from a sequence of low resolution and small view images. It's widely applied in the area such as geological reconnaissance, meteorologic surveillance, remote sensing image processing and military surveillance etc. In this paper, feature based video image mosaic is fully investigated, and a multi scale corner based video image mosaic scheme is proposed.The main work in the paper is summed up as follows:1) A multi scale Harris corner detector and descriptor is proposed in the paper. Harris corner detector is a good feature extracting method, but it is not invariant in the scale space. Moreover, it can only detect the coordinate of the corner but can't descript the corner's invariant property. We propose a multi scale Harris corner detector and descriptor which settle these problems by two means. Firstly, the corner is detected in the image multi scale space, the corner's scale information is identified, and the position is also pointed out exactly. Secondly, the corner's invariant property is descripted based on its scale information with a descriptor similar as SIFT. The experiments demonstrate that the method in the paper can repeatedly extract the multi-scale features in the images, and describe the invariant information of the features, which found a solid base for matching.2) A new feature point sets matching algorithm is proposed based on multi-scale corner's invariant gray feature and space distribution feature. In the paper, point matching problem is turned into an objective function's optimization problem, and we use the optimizing method to get the optimal matching matrix and space transformation. And the point's invariant multi-scale space feature, shape context feature, and the matching matrix constraint are taken account into the objective function. The algorithm first computer the affine transformation between the feature sets by virtual point estimation method. Then the algorithm get the matching matrix's resolution by the objective function derivative set to 0. With these two steps, the algorithm reiteratively reaching the problem's optimal resolution. The invariant feature vector and the space distribution are taken into account in the algorithm while matching, which can enhance the matching ability of the algorithm when there are noises and high-dimension space transformation.3) Video panorama mosaic is also investigated in the paper. We give out a scheme for video panorama mosaic. In the scheme, a coarse matching is applied which is based on the point's multi-scale space feature. It avoids the region of interested choosing process. Then, the fine matching scheme is carried out which proposed in the chapter 4 of the paper. Based on these, the video images are mosaicked into a panorama with planar projection model, and fused with fading out model. The proposed scheme can effectively realize the video panorama mosaic for static scenes, and provide a good technology support for military and civil application.