基于非局部自相似性的谱聚类图像去噪算法

常见的图像去噪方法只是单独地利用了无噪图像或含噪图像的先验信息,并没有将这两种图像的先验信息有效地结合起来。针对这个问题,提出一种 联合无噪图像块的先验信息和含噪图像块的非局部自相似性进行去噪的图像去噪算法。首先,对无噪图像块进行谱聚类,通过谱聚类进行学习,图像中的相似块被聚集到同一类,并将学习得到的聚类信息用于含噪图像块的聚类；然后,向量化同一类中的含噪图像块并聚集形成一个矩阵,该矩阵中包含的原始图像数据构成一个低秩矩阵；再通过一个低秩逼近过程估计出相应的原始图像数据；最后,根据逼近得到的原始图像数据重建图像。实验结果表明,相较于已有的自适应正则化的非局部均值去噪算法以及基于主成分分析和局部像素聚类的两级图像去噪算法,提出的算法不仅可以获得较大的峰值信噪比,而且还能较好地保存图像的细节,取得了更好的去噪效果。

Image Denoising Method of Spectrum Clustering Based on Non-local Similarity

The conventional image denoising algorithms just make use of the prior information of the natural image or the noise image alone,without effective combination of the prior imformation of two images to realize the image denoi-sing.For this problem,a novel image denosing method which joins the prior information of the natural image and the non-local similarity of the noise image was proposed in this paper.Firstly,the similar blocks in natural image are clustered in the same class by the spectrum clustering,and the result of the spectrum clustering with the natural image is used to get the clustering of the noise image blocks.Then,the gotten same class blocks of the noise image are vectorized as a low-rank matrix.Secondly,the low-rank approximation process is adopted on the matrix to estimate the relative original image data.Finally,the original image can be reconstructed by the estimated image data.The experimental results show that compared with the RNL(adaptive regularization of the NL-Means) and LPG-PCA(two-stage image denoising by principal component analysis with local pixel grouping),the proposed algorithm can provide significant performance improvement with respect to both PSNR and local information preservation,which produces better denoising effect.