Investigation Of Registration And Deblurring Methods For Multispectral Images

Due to the excellent ability in color reproduction,multispectral color imaging technique has shown great application value in areas such as color quality control of textile and printing,digital conservation of cultural heritage,natural spectrum analysis,computer scene rendering,material characteristics analysis,and pattern recogonition.However,the accuracy and correctness of multispectral data are extremely suffered from imaging problems in multispectral imaging system,such as misalignment between spectral images,out-of-focus blur,and motion blur.This thesis carries out innovative reasearches on aforementioned problems as follows:1.Registration metric based on gradient sparseness of difference image and registration framework are proposed to solve the misalignment problem between spectral images.Acorrding to the statistical conclusion,gradient of difference image follows a sparsest distribution when two spectral images are perfectly aligned.By reasonable definition on sparseness,the normalized total gradient of difference image metric is proposed and its validity on measuring the degree of registration is experimentally verified.A registration framework that consists of global optimization,local optimization,and image pyramid technique is proposed to solve the corresponding parametric registration problem.The experimental results show that the proposed metric and registration framework outperform state-of-the-art registration methods in multispectral image registration,and can be applied in other multimodal and unimodal registration tasks.2.A parameter self-determined multispectral image deblurring method based on interchannel correlation is proposed to solve the out-of-focus problem in multispectral imaging system with fixed focal length.The out-of-focus problem is caused by different effective focal lengths.There are three steps in the proposed method:blurriness computation,guiding image computation,and deblurring.The blurriness computation algorithm based on absolute normalized correlation coefficient is propsed by assuming the out-of-focus blur is Gaussian.According to the estimated blurriness,spectral images are divided into blurry spectral images and clear spectral images.Guiding image of blurry spectral image is linearly constructed from the clear spectral images based on a ubiquitous spectral dependence characteristic in real scenes.The characteristic is that high dimentional spectrum can be linearly approximated by rather few intrinsic spectra.The out-of-focus blur in blurry spectral image is then removed by setting the guiding image as image prior in single image deblurring framework.The regularization parameters in framework are adaptively determined by introducing and accerlerating the gerneralized cross validation.Experimental results show that the proposed multispectral image deblurring method can well remove the out-of-focus blur in multispectral images,and outperforms state-of-the-art single image based deblurring methods.3.A unified multispectral image restoration framework is proposed to solve misalignment and motion blur problems in hand-held portable multispectral imaging system.The framework consists of three steps:image registration,inpainting,and deblurring.In registration step,normalized total gradient metric is used as the objective function and direct gradient descend algorithm is used to obtain registration parameter.In inpainting step,multispectral inpainting problem is formulated as a constrained local linear model problem.A novel and rapid algorithm based on guided filtering tehchnique is proposed to solve the problem.In deblurring step,motion blur in spectral images are jointly removed by low-rank optimization.A novel algorithm called iterative singular value thresholding is proposed to solve the corresponding constrained nuclear norm minimization problem.Experimental results show that proposed unified restoration framework can be applied in restoring degraded images captured by hand-held portable multispectral imaging device.The proposed inpainting and deblurring algorithm outperform the state-of-the-art single image based inpainting and deblurring algorithms.