【作者】 刘莹；

【导师】 何小海；

【作者基本信息】 四川大学 ， 图像处理与网络通信， 2006， 硕士

【摘要】 在三维显微成像里，可以将显微镜聚焦在生物样本的不同深度来获取二维图像序列，再通过序列图像获得样本的三维信息。但是，每个聚焦层面的图像，不仅包含本层面的信息，还包含其它层面的离焦模糊信息，导致图像质量下降。通常使用两种方法来复原图像：一种是激光扫描共焦显微术(LSCM)，它可以通过光路去模糊而得到物体的三维图像；另一种是计算光学切片显微术(COSM)，它通过计算的方法来分离图像序列中焦平面上和焦平面外的光信息，从而得到样本正确的三维形态。同LSCM相比，COSM具有信噪比高，对样本不产生漂白现象和价格便宜等优势，但由于其数据量大，有运算速度慢的缺点。 本文提出了基于三维高斯点扩展函数模型的参数盲解卷积算法(PBD)。由于PBD算法需要在估计样本函数的同时估计点扩展函数(PSF)的参数，而通常采用的PSF模型较为复杂，计算量大，收敛慢，而采用三维高斯模型的PBD算法只需要估计两个参数，因此在保证复原效果的前提下，计算量大大降低了。 本文又提出了基于变化PSF模型的Hopfield神经网络图像复原算法。因为在利用Hopfield神经网络进行图像复原的算法中，都是基于PSF不变的模型，而在实际显微系统中由于样本的折射率和透镜折射率不匹配，致使不同深度的PSF是不一样的。同时为了实现三维图像序列的复原，还首次将Hopfield神经网络用于三维图像序列的复原中，并提出了基于三维高斯PSF的复原算法，取得了较好的复原效果。更多还原

【Abstract】 In the 3D microscopical imaging, 2D serial images could be recorded when the microscope is focused on different planes of the biological specimen. So the 3D information of the specimen could be recorded through these serial images. But each in-focus plane does not only have information from its own plane, but also from out-of-focus planes, which leads to the image degrading. Usually, there are two ways to restore the image. One is the Laser Scanning Confocal Microscopy (LSCM). It can get the 3D image of the object through the light route by debluring. The other one is the Computational Optical Sectioning Microscopy (COSM). It can get the correct 3D configuration of the object by separating the light information of the in-focus plane from the out-of-focus planes. Compared with the LSCM, the COSM has the advantages of high SNR, no bleaching and low price. But because of a large amount of data, it also has the disadvantage of low computation rate.In this paper, we present a Parametric Blind Deconvolution algorithm (PBD) based on 3D Gauss Point Spread Function (PSF). The PBD algorithm needs to estimate the specimen function and the parameters of the PSF simultaneously. The practical medel of the PSF is so complex that it requires a large computation and it converges slowly. But based on the simplified 3D Gauss model, the PBD algorithm only needs to estimate two parameters. While retaining the restoring effect, the更多还原