基于旋转不变稀疏表示和流形学习的图像降噪  

Image denoising via rotation invariant sparse representation and manifold learning

作　　者：汤一彬[1] 徐宁[1] 姚澄[1] 朱昌平[1] 周琳[2] 

机构地区：[1]河海大学物联网工程学院 [2]东南大学信息科学与工程学院

出　　处：《仪器仪表学报》2014年第5期1101-1108,共8页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(61101158;61201345;11274092);江苏省自然科学基金(BK20130238);中央高校基本科研业务专项资金(2012B04014)资助项目

摘　　要：提出了一种基于旋转不变稀疏表示和流形学习的图像降噪算法。首先建立基于流形学习的图像稀疏模型,使稀疏系数与图像块之间保持一致的流形结构。与此同时,由于图像块间较强的旋转相关性,算法在流形约束下采用旋转图像块进一步增强稀疏重构的性能。在具体实施过程中,则提出了一种优化算法来求解该稀疏模型。该稀疏模型经过一系列转变,使之满足经典K-SVD算法框架,并在该框架下实现迭代优化求解。实验表明,该稀疏模型,尤其对旋转图像块,能够更好地建立图像块间的结构关系,提取各含噪图像块中的隐含特征,提高重构图像的PSNR值,从而获得比传统稀疏降噪算法更高的降噪性能。An image denoising algorithm is proposed via rotation invariant sparse representation and manifold learning.The algorithm firstly builds the sparse model based on manifold learning for the image patches,and makes the spare coefficients maintain consistent manifold structure with corresponding image patches.Meanwhile,due to the strong rotation invariant relation among the image patches,the algorithm adopts block rotation to further enhance the performance of sparse reconstruction under the manifold constraint.In details,an optimal approach is proposed to solve the sparse model.With a series of model transforms,the sparse model satisfies the classic framework of K-SVD algorithm,and the iterative optimal solution is achieved under this framework.Experiment results demonstrate that the proposed sparse model can effectively estabhsh the structural relation among the image patches,especially the rotated image patches,extract the hidden features in the noisy image patches,improve the PSNR of the reconstructed image and obtain better denoising performance compared with traditional sparse denoising algorithms.

关 键 词：图像降噪 稀疏表示 流形学习 旋转不变性 

分 类 号：TN911.73[电子电信—通信与信息系统]