Research On Restoring Methods For Microscopic Images Based On CUDA And GRBF Model

During the microscopic imaging process, a series of image sequences is obtained when the microscope is focused on the different depth of a sample. Each focal plane is interfered by the scattered light because of the influence of an optical system, as a result, the observed images are blurred in the axial direction. Nowadays the confocal microscope can be used to isolate most of the scattered light through light path, yet the wide field microscope uses deconvolution methods to remove blurring in the images. Compared with confocal microscope, a wide field microscope does not yield photobleaching problems and its price is usually much cheaper. One of the research focuses in fields of microscopic image processing lies in eliminating the influence of the scattered light by the deconvolution operation.The methods of image restoration,i.e., image deconvolution can be roughly divided into two broad categories. One is the classical deconvolution method on the basis that the point-spread function(psf) of an optical system is known, which can be subdivided into no-neighbors methods, neighboring methods, linear methods, nonlinear methods, statistical methods,etc.. In the absence of priori knowledge, the other is called the blind deconvolution method which needs to identify both the psf and original image at the same time.Based on existing studies, this thesis analyses the degradation model of image at first, then introduces several restoration methods and presents a new Deconvolution and processing method for microscopic image based on continuous Gaussian Radial Basis Function(GRBF) interpolation model. Compute Unified Device Architecture(CUDA) is adopted in the implementation of GRBF method for GPU multi-threaded parallel acceleration.In the proposed method, the implementation process of deconvolution is simplified and optimized through expressing both the original image and psf by the same GRBF model. The key contents are composed of constructing the continuous GRBF model, analyzing the stability of interpolation and selecting the appropriate parameters,etc.. CUDA is adopted to accelerate parts with intensive computing such as the weighted coefficients of control points calculating and the restored image resampling. Compared with other methods, the new presented method achieves much better deconvolution results.The theory in this thesis is of a considerable reference value in the studies on deconvolution operation of continuous image model, and is expected to carry out the 3D image high-resolution deconvolution.