摘要
奇异值特征向量是图像的一种有效特征,但应用整个图像的奇异值特征向量进行人脸识别时,识别率较低。为提高识别率,提出了改进算法,对图像进行多尺度划分,先得到每个图像块的奇异值,然后将它们组合成多尺度奇异值特征向量,再应用线性鉴别分析方法进行人脸识别。由于多尺度奇异值特征向量不仅反映了整幅图像的全局特征,还反映了图像多种尺度下的局部特征,因此具有更多的鉴别信息。在ORL(Oto Rhino Laryngology)人脸库上的实验,显示人脸识别率达97.38%,且算法简单,鲁棒性更好,优于对比方法。
Singular value vector of an image is a valid feature for identification. But the recognition rate is low when only one scale singular value vector is used for face recognition. An algorithm was developed to improve the recognition rate. Many subimages are obtained when the face image is divided in different scales, with all singular values of each subimage organized and used as an eigenvector of the face image. Faces are then verified by linear discriminant analysis （LDA） under these multiscale singular value vectors. These multiscale singular value vectors include all features of an image from local to the whole, so more discriminant information for pattern recognition is obtained. Experiments were made with ORL human face image databases. The experimental results show that the method is obviously superior to the corresponding algorithms with a recognition rate of 97.38%.
出处
《清华大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2008年第10期1692-1696,共5页
Journal of Tsinghua University(Science and Technology)
基金
国家自然科学基金资助项目(60632050)
江苏省高校自然科学基金资助项目(06KJD520024)
淮安市科技发展基金资助项目(HAG07063)
关键词
人脸识别
多尺度
奇异值分解
特征组合
face recognition
decomposition （SVD）
multiscale
singular value feature combination
