Abstract:
For the other line subspace approach existing some problems of bias, overfitting and poor generalization when extracting eigenfeatures from within-class matrix, a new subspace approach is proposed. This approach decomposes the eigenfeature space into two spaces: principal component subspace and zero subspace, and regularizes the two subspaces separately to alleviate the problems of instability, overfitting or poor generalization. Experiments on ORL face base show the method achieves a given recognition rate with fewer features than other approaches and outperforms others.
Key words:
subspace method,
face recognition,
eigenfeature spectrum,
feature extraction,
recognition rate
摘要: 传统线性子空间算法在提取类内散度矩阵的特征向量时,存在偏差、过拟合和推广能力差的问题。为此,提出一种新的子空间算法。将类内散度矩阵的特征空间分解为2个子解空间,即主成分空间和零空间,再利用本征谱模型对2个空间分别进行正则化。在ORL人脸库上的实验表明,该算法使用较少的特征维数就能达到与传统算法相同的识别率。
关键词:
子空间法,
人脸识别,
本征谱,
特征提取,
识别率
CLC Number:
CENG Yue, FENG Da-Zheng-. Feature Extraction Algorithm of Face Eigenfeature Space[J]. Computer Engineering, 2011, 37(19): 148-149,152.
曾岳, 冯大政. 一种人脸本征空间的特征提取算法[J]. 计算机工程, 2011, 37(19): 148-149,152.