Adaptive Image Restoration And Learning Reconstruction Methods Based On Features

Much high-end scientific instrument like micro-focus X-ray detectors,astronomy and remote sensing imaging systems relies on low-photon-counting process.Due to the complex imaging process,there are multiple noise sources.Thus,different types of noise usually co-exist in such systems.The resulting images are influenced greatly.As one of the traditional research areas in image processing,the researches on restoration and reconstruction of images corrupted by mixed noise are showing new life with the development of such instrument.While general related work focuses on single noise model,they are not suitable for mixed noise.Moreover,most methods apply a single strategy to the whole image with no consideration of image features,which do not work well for improving the results' quality.The adaptive image restoration and learning sparse reconstruction methods can improve a lot by taking use of the local or nonlocal features of the images.They are able to maintain details as much as possible while smoothing noise,and hence promote the image restoration and reconstruction quality.Therefore,this paper concentrates on adaptive image restoration and learning reconstruction methods based on image features for mixed noise models.The main research items are as follow:(1)In the framework of regularization,the adaptive total variation image restoration objective functions based on local variance for mixed multiplicative-additive noise and mixed Gaussian-Poisson noise are proposed.Firstly,the image is segmented into different types of features based on local variance.Then,a dual adaptive scheme that both regularization operator and parameter are self-adaptive with local variance is proposed.Finally,the objective function is solved by convex optimization methods.Experimental results indicate that the proposed methods are more adaptable to different noise densities and can retain more details while smoothing noise,which will attain better results.(2)In the framework of sparse reconstruction,the learning sparse reconstruction objective function based on nonlocal information for mixed Gaussian-Poisson noise is proposed.Firstly,a clustering-based dictionary is trained,which can better match image features as well as reduce computational cost.Then,the nonlocal information is introduced as the melioration of dictionary sparse coefficients.The combination of these two is used as a new sparsity constraint.Finally,the objective function is transformed into two sub-problems under decomposition optimization and is solved via iterative re-weighted least squares minimization,which can increase arithmetic speed to some extent.The comparative experiment results demonstrate that the proposed method can improve the reconstruction quality and is more adaptable to different noise densities.