Regularization Image Restoration Methods Based On Point Spread Function Estimation

Image restoration is one of the most important research directions in the digital image processing field. Image restoration based on regularization constraint has been the hot topic in the restoration field. Based on prior knowledge of natural images, construct regularization term combined fast convergence algorithm to improve the quality of images was researched in this dissertation.In order to obtain a high quality restored image, it is necessary to estimate PSF first. Provided a precision PSF can make the recovery process faster and more stable. Because of the traditional estimation methods have higher limitations. In this paper, we will combine statistics of the natural images and maximum a posteriori model to introduce two methods to estimate PSF. One is de-convolution method and the other is Radon transform method. Both methods can estimate accurate two-dimensional PSF with any shape from special areas of the observed images, without assume the model of PSF based on the observed image. Each algorithm has a strong inhibitory effect on noise and higher accuracy estimation result.Secondly, the total variation regularization model combines TwIST algorithm for image restoration was researched in this dissertation. Each iteration involves de-noising process which is also based on total variation regularization model. TV regularization not only has a unique solution, but also can keep the image detail components during restoration and de-noising process. Regularization parameter is also very important to the quality of restored images. In fact parameter and noise variance has a special relationship, so we can match priori knowledge of the noise to calculate the parameter. During the recovery process, the parameter will be repeatedly corrected until stable. This method also can improve the quality of restored images. Taking into account the impact of noise on algorithms which are introduced in this paper, we need to estimate noise variance from the observed image to improve the performance of the algorithms and the accuracy of the results, Principal component analysis has been proved to have good results in estimated noise variance.Finally, combining the methods supposed in this paper to restore observed images, which contains different levels of noise. Objective evaluation method was chosen to evaluate the restored images, and analyzed the results in the end.