| 摘要: |
| 针对图像的非局部稀疏表示忽略图像中结构相似信息的缺点，将群稀疏表示引入到图像的最优滤波中，提出了一种有效去除图像高斯噪声的非局部群稀疏表示模型。该模型首先选择图像非局部相似块构建相似矩阵，在群稀疏限制下对相似矩阵进行正交分解得到正交矩阵；在已知噪声服从高斯分布的情况下，再通过求得的正交矩阵结合贝叶斯最小均方误差准则实现对特征矩阵的最优估计；最后通过正交矩阵与特征矩阵重构去噪后的图像。实验对比证明，所提的非局部群稀疏表示的图像去噪模型在去除噪声的同时更好地保留了图像的结构信息，获得了更好的主客观评价指标，去噪的峰值信噪比提高1 dB以上。 |
| 关键词: 图像去噪 群稀疏表示 非局部信息 贝叶斯估计 |
| DOI： |
|
| 基金项目:四川省科技计划项目（2019YJ0476,2018GZDZX0043）；人工智能四川省重点实验室开放基金项目（2016RYY02） |
|
| An image denoising model via non-local group sparse representation |
| XUE Zhishuang,YANG Pingxian,HUANG Kunchao,CHEN Mingju,CHEN Liu |
| (School of Information Engineering，Sichuan University of Science and Engineering,Zigong 643000,China;Southwest China Institute of Electronic Technology，Chengdu 610036，China) |
| Abstract: |
| To overcome the disadvantage that classical non-local sparse denoising models do not make effective use of the structural similarities of image,group sparse representation is used to modify the optimal filtering of images and a Gaussian noise denoising model based on non-local group sparse is proposed to reconstruct structural information of images effectively.Firstly,this model selects non-local patches to construct similar matrix.Within group sparse limitation,orthogonal matrix is obtained by orthogonal decomposition of similar matrix.Secondly,with the knowledge that the noise obeys the Gaussian distribution,the optimal estimation of eigenmatrix is obtained by the orthogonal matrix and the Bayesian least mean square error criterion.Finally,the denoised image is reconstructed by orthogonal matrix and eigenmatrix.The experimental results show that compared with other denoising models,the proposed model can better conserve structural information of images when denoising noise,and achieve better subjective and objective evaluation indicators,and the improvement of peak signal-to-noise ratio(PSNR) is more than 1 dB. |
| Key words: image denosing group sparse representation non-local information Bayesian estimation |
