Non-rigid Registration Of Multi-model And Multi-view Images Based On Point Features

Non-rigid registration of multi-model and multi-view images is an important research direction of image registration.It is the foundation of image fusion,multi-mode target recognition and 3-D reconstruction,and is widely used in the fields of remote sensing image processing,security monitoring,medical image diagnosis,UAV exploration,robot and automatic driving.The multi-model images obtained by different sensors vary a lot in appearance,making it difficult to extract features.The multi-view images often contain complex geometric deformation such as local projection changes,and their overlapping areas tend to be smaller.These problems add to the difficulty of obtaining robust and accurate image registration.This dissertation focuses on the research of non-rigid registration of multimodel and multi-view images.Several novel algorithms are proposed and their robustness and accuracy are verified in a variety of application scenarios.The main contents of this dissertation are as follows:1.A multi-model registration algorithm based on feature guided Gaussian mixture model is proposed to solve the degradation caused by the diversity of multi-model images.The matching of feature points turns into a probabilistic mapping problem by the introduction of Gaussian mixture model,which aims to recover more correct matches.The feature points are extracted from the marginal greyscale images to reduce the influence of multimodel diversity.A semi-supervised expectation-maximization algorithm is used to solve the model,and affine transformation is used as the global geometric constraint.The initial correspondence of the feature points is used as the anchor to aid the convergence of the model.Extensive experiments reveal that the proposed algorithm can obtain more accurate matches,and improve the accuracy and success rate of multi-model image registration.2.A neighboring features constraint based non-rigid registration algorithm of multimodel images is proposed to solve the complex non-rigid transformation problems.Referring to the idea of the kernel methods,the non-rigid transformation is modeled in the reproducing kernel Hilbert space,which convert the low-dimensional nonlinear problems to a higher-dimensional(infinite-dimensional)linear problems.Then the model can be optimized by the regularization theory.The neighboring features are used as the geometric constraint.A fast implementation based on sparse approximation is also provided to reduce the time complexity.Extensive experiments reveal that the proposed algorithm can achieve higher registration accuracy in application scenarios with complex non-rigid transformation.3.A non-rigid registration algorithm based on the composite features of edge and ORB is proposed to solve the problem of visible and thermal infrared image registration.The edge features reveal the boundaries and structural characters of the objects and thus can be used as the assistant of ORB features.A multi-feature assisted Gaussian mixture model is proposed to combine the features and recover more correct matches.The model is solved by expectation-maximization algorithm.A cross registration management procedure is introduced to assist the convergence of the model.Extensive experiments reveal that the proposed algorithm improves the success rate and registration accuracy of thermal infrared and visible image registration.4.A sparse and dense feature based non-rigid registration algorithm of multi-view images is proposed to solve the complex local deformation caused by factors like change of view.The dense feature based registration can achieve pixel-wise accuracy,but it is easily affected by rotations and scale changes.The sparse feature based registration is robust to rotations and scale changes,but cannot achieve pixel-wise accuracy.A new registration framework is constructed to combine the two features,in which they are constrained by the uniform transformation and solved by alternate iteration.Extensive experiments reveal that the proposed algorithm is robust to rotations and scale changes,and can accurately fit various local deformation.