Research On Image Denoising Via Structure Correlation

Image denoising has been studied as a fundamental problem,however,it is still a research highlight.And it is widely applied in big data mining,computer graphics,computer vision,pattern recognition and computational medicine.Due to the limitation of digital equipment,images are degraded by noise during transition and acquisition.The noise leads to the distortion of the images,and affects the credibility of the data.Furthermore,noise declines the precision of image understanding and analysis.The goal of image denoising is to surpass noise and reconstruct clean image.Image denoising algorithms can effectively improve image quality and provide reliable image data for visual dependence systems,which has important application value.Although existing methods have made certain achievements,they also ignore the importance of image structure reconstruction.The limitations of present image denoising methods are mainly as follows:1)Due to the similar properties of structure information and noise,the detail cannot be preserved.Therefore,the information in the images are incomplete.2)Image denoising models often neglect the structural consistency,resulting in the distortion of the whole structure.3)The existing color image denoising methods ignore the correlation between color channels and image self-similarity.To address these issues,the thesis carries out research oriented to image detail recovery,structure information extraction and correlation mining,and achieves the following innovative results:1.Detail preserving image denoising with patch-based structure similarity via sparse representation and SVDIn order to enhance the detail restoration quality,this paper proposes a novel denoising method that reconstructs the high and low frequency components respectively.The low frequency component represents the flat region,and high frequency component represents image structure such as edge and texture and noise.The sparse representation using patch-based structure similarity is proposed to reconstruct the high frequency parts.And the low frequency parts are reconstructed by singular value decomposition(SVD).Finally an energy minimization function that contains high and low frequency parts are presented.The experimental results show that the numerical improvement is 0.2-0.84 db on average compared with the then frontier and classical algorithms.2.A competent image denoising method based on structural information extractionGenerally,the existing algorithms ignore the structural consistency.Therefore,the thesis propose a structure extraction model to be a fidelity term of the image denoising model.The structure information extraction algorithm is verified on noisy images,which effectively perceive structure under noise.Also,a competent image denoising algorithm is designed by the proposed model and image frequency features.This algorithm further enhances the quality of restored image.The proposed method is validated on the Setl2 and BSD68 datasets.And the experimental results show that this method achieves numerically competitive results and significantly outperforms the state-of-art methods in terms of visual quality.3.Color image denoising based on self-correlationThis method offers a novel color image denoising method that applies quaternion to the group sparsity model,where each pixel is expressed as a pure quaternion.At first,each pixel of the observed image is expressed as a quaternion unit.And the quaternion patch group matrix is established by Pearson’s correlation coefficient.After this is set,the proposed model learns the dictionary for each patch group,working well with the pursuit algorithm.In other words,the group-based sparsity method assumes that each patch group is a linear combination of the basic elements of the dictionary.Unfortunately,it is still arduous to reconstruct the image structure precisely.Therefore,the group sparsity model incorporates kernel Wiener filtering to enhance image structure quality further.Fueled by the exploration of the inner correlation of color channels,the proposed method can preserve the image information as much as possible while removing noise.The experiments validate the efficiency of the proposed method both in numerical results and visual performance on different noise levels.In summary,the paper addresses the problems of poor image detail recovery,lack of structural consistency constraints in denoising models and color image denoising algorithms that do not consider the correlation between color channels and image self-similarity,and innovatively proposes a series of image denoising algorithms to improve the accuracy of image understanding and analysis.