Study Of Deconvolution Methods Of Fluorescence Microscopic Images

In fluorescence microscopic imaging technique, ultraviolet rays are taken as irradiation source and illuminate the testing cells. After this the cell structures are bioluminescent and then people can observe the cell appearance and other features under a microscope. For example, the vivo structure and activity mechanism of biological cells can be observed by Fluorescence microscopic imaging technique in biomedical research. Because of the interference of many factors in imaging system, the image resolution will decrease. In general two kinds of deconvolution methods: optic and calculative can be used. The scattered light is blocked before it receive the detector in the optic methods and the distortions are decreased. Although the three-dimensional resolution of optically treated images is more obvious, the character of anisotropic is serious. In order to reduce the noise and blur in imaging results, the data is processed by computer in calculative deconvolution and we used this method to improve the image resolution in this paper.Three ideas are offered in this paper. The traditional deconvolution methods are usually non-blind kind and can be used when the point spread functions are known. We firstly designed a deconvolution method which was based on the model of the parameterized point spread function. Then the point spread function was fitting by multiple Gaussian functions and combined it with the method of ER deconvolution. This blind form of ER deconvolution could be solved by the method of alternative minimization. We also applied the penalties of different orders to our problem of ER deconvolution. The second idea was that we presented the non-local regularization and the likelihood term remained unchanged. Then we applied it to two-dimensional space images and three-dimensional images of time series and solved it with steepest descent method. The third idea took the character that the lateral axis resolution was high than vertical axis resolution in three-dimensional fluorescence microscopic images. We interpolated the vertical axis and made the lateral axis resolution is in accordance with the resolution of vertical axis. And then we learned the dictionary from lateral axis and applied it to vertical axis by the method of sparse representation. Combined with the non-local prior of vertical axis, the prior of sparse representation greatly improved the resolution of vertical axis.This paper conducts computer simulationed with MATLAB 2012 and Visual Studio 2012. We compared the results of the three methods with Richard-Lucy algorithm, RichardLucy + TV algorithm and wiener filtering. Then the deconvolution results were compared and analysed with two performance indexes: Power Signal-to-Noise Ratio(PSNR) and Structural Similarity Index(SSIM). Comparing the results of simulation data and real data, we found that our three methods had remarkable deconvolution results.