摘要
主成分分析(PCA)已经广泛应用于计算机视觉中,但是传统的基于L2范数的PCA对异常值和特征噪声(比如有遮挡的图像)敏感.基于L1范数的PCA(L1-PCA)相比基于L2范数的PCA更具鲁棒性,并且可以克服对异常值和特征噪声敏感的问题.然而,在应用L1-PCA算法时,其算法的优化非常关键.本文针对这一问题,提出基于增强拉格朗日乘子的L1-PCA的优化算法并将其应用于处理有遮挡图像的重构,通过在Yale人脸数据库的实验测试表明所提出的算法有效.数值和可视化的实验结果都表明优化的L1-PCA优于传统PCA.
Principal component analysis（PCA） is widely used in computer vision,therefore traditional PCA based on L2 norm is sensitive to both outliers and feature noises（such as occlusions of images）.Compared with L2-norm PCA,L1-norm Principal component analysis（L1-PCA）has better robustness and overcomes sensitive questions about outliers and feature noises.However,the optimization algorithm for L1-PCA application is very important.In order to solve this problem,the paper proposes the algorithm based on Augmented Lagrange Multiplier and applies it in reconstructed image against image occlusions.Extensive experiments on Yale face data sets verify the efficiency of the proposed algorithm.Both numerical and visual results show that the optimized L1-PCA is better than standard PCA.
出处
《小型微型计算机系统》
CSCD
北大核心
2013年第1期173-176,共4页
Journal of Chinese Computer Systems
基金
2012年国家科技支撑计划项目(2012BAH08B00)资助
湖南省自然科学基金项目(06JJ50143)资助
关键词
主成分分析
L1-PCA
增强拉格朗日乘子
图像重构
principal component analysis
L1-PCA
augmented lagrange multiplier
reconstructed image
