Research And Implementation Of Adaptive Methods For Regularization Based Image Restoration

Image restoration is one of the most important research directions in the field ofdigital image processing. Regularization method is a good choice to solve the ill-posedrestoration problems. Appropriate regularization term and parameter determine theefficiency of the algorithm. In this article, we proposed two methods in adaptivelyconstructing regularization term and selecting the regularization parameter to use thecorrect priori knowledge. In this way, the efficiency of the algorithms is improved.First, based on the Total Variation model, we proposed an adaptive regularizationparameter estimating method. This model assumes a Laplacian distribution of naturalimage in gradients. In addition, we assume that noise is of the Gaussian distribution.Using the maximum a posterior (MAP) method, this article formulates the way toestimate the regularization parameter. Our method is easy to operate, and experimentsshow that the performance of the algorithm improved much using our adaptive method.Then, based on hyper-Laplacian prior, we present a method adaptively selectingregularization parameters and norm, appropriate values of which correspond to correctprior knowledge of image and noise, and determine the performance of the algorithm.Therefore, we proposed a MAP method to adaptively select the two values, using thedistributions of image and noise. With MAP, we jointly estimate the latent image,regularization parameter, and norm. Experiments show that our adaptive method canimprove restoration results compared with fixed method.Finally, using VC++, we systematically validate our image algorithms. The systemimplements two algorithms: our adaptive algorithm and the fixed algorithm. In the fixedalgorithm, norm values can be set for users to observe and analyze the impact of normsfor restoration tasks. Our system consists of two parts: the control module and resultdisplay module. In the control module, the user can select the blurring template, inputnoise level and choose algorithm for run; the result display module will display the clear,degraded, distribution of the image in gradients and restored image, so that the users canvisually compare the status of the image under different conditions. Using our systemrestoration, the adaptive algorithm and the fixed algorithm are compared.