Studies On Image Mosaic

Image mosaic is one of the hot topics of image based rendering (IBR) these years. In this thesis, some efforts on it are performed and, specifically, the following improvements and new methods are presented:(1) A Novel Feature-Based Image Mosaic Method In this method, the features are first extracted and described by Scale Invariant Feature Transform and the approximate-nearest neighbors of each feature are found out by priority search. Then the homography is calculated by RANSAC, and finally a bundle-adjustment algorithm is applied when there are more than two source images. An improved RANSAC algorithm based on the modified median flow filter is presented to improve the stability and accuracy of homography calculation. A bundle adjustment algorithm based on the minimization of the distances of feature matches is also introduced to remove the accumulation of errors.(2) An Improved Image Mosaic Method Based on Dynamic Programming This method first calculates phase correlation and then reaches image mosaic by dynamic programming and mulitresolution spline mosaic. Dynamic programming is used to find a best seam-line for deghosting while multiresolution spline mosaic is used to realize mosaic without exposure difference. A new gradient operator emphasizing more the geometry similarity than former operators and a simple and efficient search rule based on this operator are proposed in order to find a robust seam-line in a convenient way. Hierarchical multiresolution spline mosaic is also introduced to improve the efficiency of exposure difference removal.(3) A Novel Image Mosaic Method Based on Graph Cut This method first calculates phase correlation and then reaches image mosaic by graph cut and poisson image fusion. Graph cut is used to find a globally optimized seam-line for deghosting while poisson image fusion is used to realize smooth mosaic without exposure difference. A new weight calculation method based on the statistics of gradient direction histogram is presented here in order to robustly realize graph cut. An edge transition poisson image fusion method is also proposed to reach exposure difference removal.(4) Manifold Mosaic for Large Displacement Images This method first cuts the overlapping area into multiple strips and then warps them together interleavingly. Finally these warped strips are blended with multiresolution spline mosaic method. In order to improve the efficiency of multiresolution spline mosaic, a method called self-copy multiresolution spline mosaic is presented in this thesis. Exposure differences can be eliminated and deghosting can also be fulfilled to some degree with this manifold mosaicmethod.Based on the methods described above, applications on image mosaic are also explored in this paper and several prototype systems are introduced: Easy Panorama, Easy Tour and Virtual Dachang. EasyPanorama fulfils the whole authoring process of panorama: from the reading of source images to the publishing of the panorama. Easy Tour integrates panorama and Tour Into the Picture, and supports virtual tour of large area. Virtual Dachang integrates panorama, Tour Into the Picture and geometrical modeling. It can fulfil digital protection and exhibit of the culture heritage of Dachang.