Super-resolution Image Reconstruction Based on Mutual Information Registration
|School||South China University of Technology|
|Course||Signal and Information Processing|
|Keywords||Super-resolution Image Reconstruction Mutual Information Registration Projections onto convex sets Binary Tree Complex Wavelet Transform|
The super-resolution image reconstruction technique is by determining the relationship of the motion between the low-resolution observation images from multiple frames of the same scene, that they contain a variety of complementary information and the redundant information is fused to the single-frame resolution image, and to eliminate blurring and noise in the image signal processing technology, is a time efficiency in exchange for the spatial resolution of the signal processing method. It can solve the problem of costs, hardware conditions or imaging environment factors not directly from the hardware meet the requirements of high-resolution images. The purpose of this paper is to study by the super-resolution technology, the establishment of a complete and effective method of reconstruction observation image details. This paper first analyzes the basis of super-resolution reconstruction problems of mathematical physics, and then study the observed image of the imaging process, the establishment of a complete imaging system degradation model, established the mathematical equations that describe the observed image with the original image signal between the and in accordance with mathematical relationship, the more in-depth description and analysis of the super-resolution problem. Then, two important aspects of the super-resolution reconstruction, image registration and image reconstruction, respectively start the study. Image registration is the process determine the relative positional relationship of the plurality of observation images, the high-precision image registration is the basis of the accurate reconstruction of the image details. In image registration, this paper studies the broader application of the mutual information registration method, this method is especially suitable for multi-modality image registration. Less than for traditional mutual information registration, this article by dual-tree complex wavelet transform combined with edge information to improve mutual information measure, at the same time, the registration PV interpolation method extended and improved in order to reduce mutual information into the possibility of local extrema. The entire mutual information registration method is carried out on a binary tree complex wavelet transform pyramid model. Reconstruction in image resolution, first proposed image interpolation method based on binary tree complex wavelet transform, and the method is applied behind the focus on the two proposed super-resolution reconstruction methods: combined binary tree complex wavelet transform interpolation fusion image fusion method and improved the POCS image reconstruction method. In the POCS's improvements, mainly to improve the image reconstruction process initial value; adaptive relaxation projection operator to accelerate the convergence rate of the POCS algorithm; parallel projection framework to improve the the POCS algorithm may not converge in situation; the same time, the introduction of energy non-decreasing constraints and smoothness constraints to further ensure the quality of the reconstructed image to suppress the ringing effects of the reconstruction process. Based on the above work, the paper established a relatively complete super-resolution reconstruction process. Experiments show that this paper, mutual information registration method and image reconstruction methods can be more effective recovery of the details of the imaging target.