低秩张量填充及降噪模型的彩色图像复原方法

Color Image Restoration Method Based on Low-Rank Tensor Filling and Noise Reduction Model

现有大多数彩色图像去噪方法分开处理输入图像的彩色通道,利用张量挖掘通道间的相关性,补全带有缺失的彩色图像以及图像去噪.从少部分观测项中恢复一个低秩张量,并且对存在于图像中的各类噪声针对性的去除.对广泛分布于多维成像数据中的非局部自相似性以及稀疏线性逼近中的相似斑块进行分组,通过使用即插即用框架将两者结合.文章提出一种适用于即插即用框架下用块对角表示的张量填充去噪方法,使用去噪算法作为基于模型反演的先验.将全局截断奇异值分解与局部鲁棒主成分分析结合,能够利用更多空间信息,附带不完整信息且含噪声的图像能够完整复原.实验显示使用块对角去噪方式比其他去噪方式在峰值信噪比及结构相似指数上皆有提升,证明该方法是一种精确度较高的方法.

Most of the existing color image denoising methods deal with the color channels of the input image separately.In this paper,we use tensor to mine the correlation between channels to complete the missing of color images and denoise images.A low rank tensor is recovered from a small number of observation items,and all kinds of noises existing in the image can be specifically removed.Non-local self-similarity widely distributed in multi-dimensional imaging data and similar patches in sparse linear approximation are grouped.Plug and Play framework is used to combine the two.In this paper,we propose a tensor filled and denoising model which is represented by block diagonals under the Plug and Play framework,and use the denoising algorithm as a priori of model-based inversion.By combining global truncated singular value decomposition with local robust principal component analysis,images with incomplete information and noise can be completely restored using more spatial information.The experiment shows that the block diagonal denoising method has higher pick signal to noise ratio and structure similarity index than other denoising methods,which proves that the method is more state-of-the-arts.