A semi-supervised learning algorithm based on regularization on graphs is presented and is applied to recognition and retrieval of face images. In a learning phase, the value of classification function is fixed at labeled data and that of unlabeled data is estimated by a regularization scheme whose solution is computed with iteration methods. In a classification phase, the value of classification function of a new datum is computed directly from those of learning data without iterations. The classification rate of the present method is higher than that of the conventional methods such as the basic nearest neighbor rule and the eigenface method. Similarity search of data is also a particular case of the semi-supervised learning where a query is labeled and all data in a database are unlabeled. The relevance degree of data in the database is calculated with regularization and some data with high relevance degree are outputted. The precision of this retrieval scheme is higher than that of the basic similarity search methods.
[1]
Fan Chung,et al.
Spectral Graph Theory
,
1996
.
[2]
Yixin Chen,et al.
Content-based image retrieval by clustering
,
2003,
MIR '03.
[3]
Zoubin Ghahramani,et al.
Combining active learning and semi-supervised learning using Gaussian fields and harmonic functions
,
2003,
ICML 2003.
[4]
Bernhard Schölkopf,et al.
Learning with Local and Global Consistency
,
2003,
NIPS.
[5]
William I. Grosky,et al.
Negotiating the semantic gap: from feature maps to semantic landscapes
,
2002,
Pattern Recognit..
[6]
J. Ortega.
Numerical Analysis: A Second Course
,
1974
.