Study On Fast Image Registration Algorithms In Pattern Recognition

The study of image registration algorithm is one of the branches in the field of Pattern recognition. The core of the problem is to find the best mapping function between two images where one image is a transformed version of the other. This mapping function needs to make the similar part of the two images overlapped as much as possible. The efficiency and accuracy of image registration algorithm plays an important role in the final results in many fields. Since image registration has an extensive application area with widely varied data sources, so far different methods are somewhat limited only to certain application problems.Among the image acquisition approaches, one of them is to acquire a series of projecting data at first, then we could recover the image of the inner part of the objects through these data. Both CT(Computerized Tomography) and PET(Positron Emission Tomography) in modern diagnostic medical imaging obtain the image of inner human body through similar methods. Registration of this kind of images are accomplished in the recovered time domain images. This thesis focuses on the study of the nature and realization of some common projecting models, such as RT(Radon Transform), DBT(Divergent Beam Transform) and HT(Hough Transform) which is frequently used in the pattern recognition research. The work in this thesis also includes the analysis of the switching relationships between them and makes a summary of fast algorithms for RT. On the basis of these, the thesis puts forward a new idea for registrating the images obtained from projection data. The idea is registrating the images only from their projections. In this way, we can avoid registrating images which already contain some errors when recovered from projection. So far no results have been obtained that solves the recognition problem completely in the projection domain due to coupling of transform parameters. We develop a novel algorithm based on this new idea for such parameter decoupling in the projection domain. The algorithm can accomplish the registration of two images which have affine tranform relationship through using FFT and least squares method on their projecting data. Simulation shows that the proposed algorithm has a high computational efficiency and is robust to noise. Compared with the state-of-the-art approaches such as [113], the proposed algorithm has a big improvement, in computational efficiency. The algorithm is also suitable for the registration of ordinary images. It also has some advantages in computation efficiency and accuracy compared to the image registration algorithms based on FFT. For the accuracy of the parameter estimation in the proposed fast algorithm, the thesis also puts forward an improved Phase Correlation method. The method overcomes the defect of the traditional method which has a big error in low-frequency signal registration. The improved Phase Correlation method achieves quite good registration effects in the proposed fast algorithm.This thesis systematically makes a summery of the classical image registration algorithms from the point of view of the image global gray value and the image local feature. We also analyze the advantages and disadvantages of various algorithms. And an introduction of currently extensively used SIFT(Scale Invariant Feature Transform) method which is based on the image local feature points and a fast registration algorithm which is based on the first-order moment of the images is made.