Research On Pixel-level Image Fusion And Its Related Technologies

With the development of thirty years, the multisource sensors information fusionis gradually becoming a rising discipline. The multisource sensors image informationfusion need to integrate the images information of the same scene captured by two ormore sensors into a single image, which can describe the scene more accurately, reliablyand comprehensively. With the further development and improvement of image fusiontechnologies and its related theories, it can foresee that image fusion will be more widelyapplied to the fields of military, medical, industrial monitoring, remote sensing of theearth, and so on. Although great achievements have been made in the field of imagefusion, it still has many new situations and new issues to face for the image fusion, whichmakes image fusion become more important. At present, the domestic research on theimage fusion is just getting started, but is far behind the foreign counterpart. Therefore,it is necessary to carried out in-depth study on the image fusion. In this dissertation thefollowing works are focused on the researches of theories and technical issues of thepixel-level image fusion:(1) The image fusion is an ill-posed inverse problem, and the use of the simulatedannealing algorithm for solving the energy minimization function is very slow, in thesame time, the obtain of the optimal solution for the problem can not be guaranteed,in this dissertation graph theory is applied to construct the corresponding graph modelfor energy minimization function of the image fusion, and the graph cuts theory is usedto optimize the solution, on one hand the speed of solving the image fusion problem isgreatly improved, on the other hand, the global optimal solution can also be obtained.(2) Based on the subspace and the multi-scale method, extensive research on imagefusion have been carried out. First, the two-dimensional principal component analysis andsteerable pyramid decomposition is applied to fuse the multi-spectral and panchromaticimages, in addition, the algorithm about the edge protection is also taken into account.The simulation indicated that the proposed methods can effectively improve the spatialresolution and reduce the distortion of the spectral information; Second, the principalcomponent analysis, the IHS transform and visual driving model are comprehensivelyused to study the medicine images fusion of the PET and MRI images. The virtues of the three algorithm are combined and can improve the spatial resolution and reduce thedistortion of the spectral information. Third, a new independent component analysis al-gorithm(ICA) is proposed based on the special linear group theory. This algorithm canimprove the fusion effect greatly. At last, a new fusion method is proposed based onthe maximum likelihood estimation theory and the Laplacian pyramid decomposition al-gorithm, the algorithm effectively combines the virtues of spatial estimation theory andthe multi-scale decomposition, the experimental results show that the algorithm is able toobtain better fusion performance.(3) For the noisy source image fusion, in order to improve the spatial resolutionand visual effect, and to protect the edge information, a modified total variation model isproposed, combined with the second-order optimization model, a new fusion algorithm isalso proposed.(4) Because there are many matrices eigenvalue solution problems in the image fu-sion, and neural network is parallel, fast, and easy to be realized in the matrices eigenvaluesolution, this dissertation systematically studied the eigenvalue problems by neural net-works, and the followings can be obtained: all of the eigenvalues and the correspondingeigenvectors of real antisymmetric matrices, all of the eigenvalues and the correspondingeigenvectors of special orthogonal matrices, eigenvalues and the corresponding eigen-vectors of imaginary part absolute largest or smallest in the eigenvalues of general realmatrices, eigenvalues and the corresponding eigenvectors of modulus largest or smallestin the eigenvalues of general real matrices, eigenvalues and the corresponding eigenvec-tors of real part largest or smallest in the eigenvalues of general real matrices, and the realpart of eigenvalues with real part absolute largest in the eigenvalues of general real matri-ces. In the same time, a complex neural networks algorithm is also discussed to solve theeigenvalues of real antisymmetric matrices, and a unified model is proposed to obtain alleigenvalues of general real matrices.