| This research is concentrated on the development and expansion of the maximum likelihood blind deconvolution (MLBD) algorithm for three-dimensional light microscopy. Based on a Poisson random point process model and the maximum likelihood estimation of unknown signal, this algorithm is used to remove the unwanted blur and to estimate the undistorted fluorescence dye concentration. The significance of this algorithm is that it does not require prior knowledge of the Point Spread Function (PSF) of the microscope. Instead, an estimate of the PSF is produced along with the deconvolved image data.; In this thesis, the MLBD algorithm is improved for robustness, extended for more applications, and expanded to achieve sub-pixel resolution. The major topics and contributions are: (1) Data correction and pre-processing: Efficient data correction algorithms are developed to correct for non-uniform photosensitivity, bad pixels, bias level of the CCD camera, intensity flicker in the optical slices, and intensity attenuation in the image stack. We also evaluate how these data correction schemes affect and improve the deconvolution results. (2) The effects of first guesses, constraints, and regularization: By using normalized lag-likelihood and I-divergence as the quantitative measurements, we systematically investigate how PSF constraints and various forms of image and PSF first guesses affect the MLBD process. A novel PSF intensity regularization scheme is also derived to prevent the delta function PSF and over-fitting phenomenon from occurring. (3) Blind deconvolution for different microscope modalities and validation of deconvolution results: We develop effective strategies to further extend the MLBE algorithm for more microscope modalities, including the spinning-disk scanning confocal and the multi-photon excitation microscopes. This thesis also evaluates the performance and accuracy of the MLBD algorithm by analyzing the reconstructed object and PSF of known object qualitatively and quantitatively. (4) Sub-pixel resolution enhancement: A novel sub-pixel resolution blind deconvolution algorithm is developed to achieve sub-pixel resolution improvement. This sub-pixel deconvolution algorithm is based on the MLBD algorithm. This new algorithm is tested by computer simulations and processing fabricated test targets and biological specimens. |
