NSCT And Graphical Models With Application In Image Registration

Image registration is to overlay two or more images of the same scene taken at different times, from different viewpoints or by different sensors. It is an important and foundational issue in pattern recognition and computer vision. It is also a prerequisite step prior to image fusion or image mosaic. Its research result is widely used in many areas, such as remote sensing, medical imaging and multi-sensor fusion based target recognition.Image registration can be roughly divided into intensity-based registration and feature-based registration. As feature-based registration has very prominent advantages in terms of reliability and robustness, the registration methods which are feature-based become popular recently. These methods extract local invariant descriptors as image features. and then, image registration is equivalent to finding correspondences between two feature sets. But Image features are diversity and variability, by now, Feature-based registration is still an open problem, which is a hot topic many researchers focus on.NSCT and graphical models with application in image registration are studied in this thesis. We switch image matching and image registration from the spatial domain into the NSCT domain and improve the performance of image registration with SIFT algorithm and graph model. We propose three algorithms in this thesis, the main research contents and the achievements are as follows:1、A remote sensing image registration algorithm based on SIFT and NSCT is proposed in this thesis. It combines the flexibility of NSCT for image decomposition with the effectiveness of SIFT algorithm for feature representation to register remote sensing images. Firstly, two related images are decomposed by NSCT; Secondly, the low frequency images are inputted to SIFT algorithm to obtain matching results; Finally, the optimal parameters of transformation model are estimated based on the matching results, re-sampling and bilinear interpolation are employed to complete the registration for the two related images. The experimental results demonstrate the better effectiveness and the higher matching accuracy of the proposed algorithm by comparing with the other two algorithms.2、A spectral matching algorithm based on cycle basis is proposed in this thesis. Firstly, according to the feature points of two related images, two groups of cycle basis are found respectively; Secondly, construct the weighted adjacent matrices from two groups of cycle basis respectively and submit the matrices to singular-value decomposition(SVD); Thirdly, complete matching matrix and initial probability matching matrix from the result of SVD; Finally, acquire the matching result by using doubly stochastic matrix. Experimental results demonstrate the feasibility and effectiveness of the approach.3、A new spectral matching algorithm is proposed by using the nonsubsampled Contourlet transform and SIFT feature. The nonsubsampled Contourlet transform is used to decompose an image into a low frequency image and several high frequency images, and the scale-invariant feature transform is employed to extract feature points from the low frequency image. A proximity matrix is constructed for the feature points of two related images. By singular value decomposition of the proximity matrix, we obtain a matching matrix (or matching result) reflecting the matching degree among feature points. Experimental results indicate that the proposed algorithm can reduce time complexity and possess higher accuracy.