荧光显微图像的解卷积方法研究

发布时间：2018-03-24 12:34

本文选题：荧光显微　切入点：点扩散函数　出处：《华中科技大学》2016年硕士论文

【摘要】：荧光显微成像方法是采用紫外线作为照射源,来照射被检测细胞,使细胞结构发出荧光,然后在显微镜下观察细胞的形状及其他特征。在生物医学研究中,通过荧光成像技术,可以查看到生物细胞的活体结构和活动机理。但是由于各种因素的干扰,成像系统得到的图像分辨率都会降低,而解卷积的目的就是提高退化后图像的分辨率。通常有两种可用的解卷积方法:光学的和计算的。光学方法是在散焦光到达检测器之前就阻挡它以减少失真。虽然光学处理的图像的三维分辨率更为明显,但是该图像包含严重的各向异性。在计算解卷积的方法中,数据则是由计算机处理,以减弱图像中的噪声和模糊,本文就是通过计算的方法来提升图像分辨率。本文提供了三种解卷积的思路。传统解卷积方法均为非盲解卷积,仅适用于点扩散函数已知的情形,我们设计了一种基于参数化点扩散函数模型的解卷积方法,将点扩散函数拟合为多重高斯叠加形式,构造出了ER方法的盲解卷积形式,然后采用轮流优化方法进行求解,同时我们将结合TV与Hessian正则的多重正则方法用于ER解卷积。第二种思路是在似然项不变的情况下提出了Non-local正则项,并应用到二维空间图像与三维时间序列图像中,并采用最速下降法求解。思路三则是利用三维荧光图像纵轴分辨率低于横轴的特点,将图像沿纵轴进行插值,以使与横轴分辨率一致,然后采用稀疏表示的方法从横轴学习字典应用到纵轴,同时结合纵轴的Non-loal先验来提高纵轴的分辨率。本文采用MATLAB R2012a和Visual Studio 2012软件进行计算机仿真,并将每种方法的实验结果都与Richard-Lucy、Richard-Lucy+TV和维纳滤波方法进行了对比,并对解卷积结果针对峰值信噪比(PSNR)、结构相似度(SSIM)两个性能指标进行了比较和分析。通过对仿真数据和真实数据的对比实验发现:基于参数化点扩散函数的盲解卷积模型、基于Non-local正则项的解卷积模型与基于稀疏表示的解卷积模型这三种思路都有着较为显著的解卷积效果。
[Abstract]:Fluorescence microscopic imaging method is used as the ultraviolet irradiation source, irradiated cells were detected, the cell structure of the fluorescence and then observe the cell shape and other characteristics under the microscope. In biomedical research, through the fluorescence imaging technology, you can view the structure and biological activities in vivo cellular mechanism. But due to various factors the imaging system, the image resolution will be reduced, and the deconvolution is to improve resolution of image degradation. There are usually two deconvolution methods available: optical and optical method is calculated. The defocus before the light reaches the detector will block it to reduce distortion. Although 3D optical resolution image processing the more obvious, but the image contains serious anisotropy. In the calculation method of deconvolution, data is processed by a computer, to reduce the noise in the image And fuzzy, the work of this paper is to improve the image resolution by calculation method. Provides three deconvolution ideas in this paper. The traditional deconvolution methods are non blind deconvolution, applies only to the point spread function is known, we designed a deconvolution method based on parametric point spread function model, the the point spread function fitting for multiple Gauss superposition constructed form ER method for blind deconvolution, and then solved by alternate optimization method, we combine multiple regularization method TV and Hessian regular ER for deconvolution. Second kinds of ideas is the likelihood unchanged under the condition of Non-local is put forward and the regularization term. Applied to 2D space and 3D image sequence images, and uses the steepest descent method. The ideas of the three is to use the 3D fluorescence image below horizontal longitudinal resolution characteristics of the image along the longitudinal axis In order to make the interpolation, and horizontal resolution, and then using the sparse representation dictionary is applied to learning from the horizontal axis, and combining with the longitudinal axis of the Non-loal prior to improve the longitudinal resolution. This paper uses MATLAB R2012a and Visual Studio 2012 software for computer simulation, and the experimental results of each method and Richard-Lucy, the comparison of Richard-Lucy+TV and Wiener filtering method, and the deconvolution results in peak signal-to-noise ratio (PSNR), the structural similarity (SSIM) two performance indexes were compared and analyzed. Based on the simulation data and real data experiments found that parameters of the point spread function of the blind deconvolution model based on convolution model Non-local regularization based on sparse representation based deconvolution model of the three methods have obvious deconvolution effect.

【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】