Image Restoration Algorithm Base On Phase Diversity For Ground-based Telescope And High Speed Implementation On GPU

The imaging resolution of ground-based telescope is seriously limited by theeffects of atmosphere turbulence and optical system aberrations. Especially is theatmosphere turbulence induced distortion more serious with larger aperture thatmakes the performance of imaging degraded, with the results that the detectionability of large ground-based telescope is restricted. Phase Diversity (PD) techniquethat can overcome wavefront distortions becomes an important developmentaldirection of image restoration for ground-based telescope. By making use ofmaximum likelihood estimate to construct iterative optimization model, it jointlyestimates image of object and wavefront phase from simultaneous collection ofpairs of short exposure image with known phase diversity. It is a highperformance-price ratio method of simple configuration, low cost and well fit forboth point sources and extended sources.In this paper, referred to research advances in domestic and overseas, a largenumber of numerical simulations and experiments are carried out, consequently,PD algorithm is discussed and analyzed in depth. Besides, the performance of itshigh speed implementation on GPU parallel processing platform is verified. Themain works and achievements are as follows:1. After analysis for the error factors found in PD test bed alignment and data processing procedure, mathematical model of PD is improved. An adaptive filterbased on Butterworth low pass filter is suggested to reduce noise effect, filter therestored image and at the same time improve convergence properties, so as to reducethe overall computational time. Meanwhile, a two-step auto registration method toimprove the reliability and accuracy, by use of phase correlation combined with leastsquare surface fitting, is proposed.2. It is difficult to achieve real time application of PD on CPU, and DSP andFPGA is a proper way to improve its performance. While the complex structure ofthe PD objective function influences on its hardware implementation. According tothe theory of Zernike polynomial, a method to modify the PD objective function isproposed, by which the computation of the PD objective function only depends onthe polynomial and the hardware implementation of DSP, FPGA and parallelism ofhardware processing are more easily.3. PD algorithm is suitable for parallel computing based on GPU, by reasons ofits data partitioning features. Thus a parallel algorithm of PD based on theheterogeneous architecture consisted of CPU and GPU is proposed. According to theanalysis for PD algorithm based on CPU, using the programmer friendly CUDAframe, the parallelization on pixel level is achieved. The construction of objectivefunction and its derivative, the most time consuming part, is transplanted to the GPUplatform, while the iterative process is still controlled by CPU. Accordingly, thespeed of PD image restoration can be greatly increased by taking full advantages ofGPU parallel computing power and floating point calculations.4. A series of PD experiments are implemented on the high speed computingsystem, including restoration of spread target near ground and spatial objects. Theexperimental results show that PD image restoration technique can significantlyimprove the imaging quality of ground-based telescope imaging system. It makes asingle invisible star to be detected and full width half maximum decrease by50%,the undistinguishable binary star with a separation of0.4″is successfully resolvedfrom speckle noise, and details of spread target which are too fuzzy to be identified are improved to the level of identifiable. Meanwhile,under the premise of the qualityof image can be guaranteed, the CUDA implementation can obtain speedup of up totwo orders of magnitude over the CPU serial implementation. The real time abilityof PD image restoration procedure is effectively improved that makes the PDtechnique more practical in the telescope imaging system.