基于预学习字典的对称ADMM在人脸去噪上的应用  

Application of Symmetric ADMM Based on Pre-Learned Dictionary in Face Denoising

作　　者：张静雯 党亚峥[1] 倪诗皓 

机构地区：[1]上海理工大学管理学院,上海

出　　处：《建模与仿真》2024年第2期1630-1640,共11页Modeling and Simulation

摘　　要：字典学习算法是图像去噪的常用方法,但字典训练是一个耗时的过程,且会丢失图像中的有效信息。本文提出一种基于预学习字典的低秩算法,将图像中的干净数据和结构化噪声用对应的字典和系数矩阵表示。首先通过消融实验证明模型各部分的作用,接着对部分损坏的图像进行训练并与其他算法对比,最后将算法训练出的映射矩阵用于测试数据。结果表明PLID对于噪声的处理优于其他算法,测试数据经过映射矩阵的线性变换后也可以去除噪声。Dictionary learning algorithm is a common method for image denoising, but dictionary training is a time-consuming process and can lose valid information in images. In this paper, a low-rank algorithm based on pre-learned dictionaries is proposed, which represents the clean data and structured noise in the image with the corresponding dictionaries and coefficient matrices. Firstly, the role of each part of the model is proved by ablation study, then the partially damaged images are trained and compared with other algorithms. Finally, the mapping matrix trained by the algorithm is used for the test data. The results show that PLID can deal with the noise better than other algorithms, and the test data can also remove the noise after the linear transformation of the mapping matrix.

关 键 词：预学习字典 对称ADMM 人脸去噪 消融实验 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]