Metric Learning For Face Analysis

Face analysis has always been one of the most popular research topics in computer vision and pattern recognition.Face analysis attempts to obtain the identity,age,gender,expression and other key information of a person by taking advantage of face images.Face analysis tasks include face recognition,kinship verification,etc.There are two major components of face analysis:1)the descriptors employed to represent the images;2)the metric applied to compare the descriptors.However,traditional metric methods often can not achieve satisfactory results,without the usage of prior knowledge of training dataset,and thus these methods can not effectively reflect the relationship between the samples.Therefore,it is of great significance to apply metric learning to face analysis.The main works of this thesis are organized as follows.Firstly,we extensively investigate and summarize the existing imetric learning algo-rithms,and introduce the principles and representative algorithms of the three types of metric learning techniques,which are supervised metric learning,unsupervised metric learning and semi-supervised metric learning.respectively.Secondly.traditional metric learning algorithms usually use the fixed bounds to constrain the distance between samples of pairs and learn a high rank metric matrix from the given training data.However.the learned metric may not perform well using the fixed bounds and a high-dimensional matrix.In this thesis.we propose an adaptive metric learning with the low rank constraint method.which restricts the learned distances betwen samples of pairs using adaptive bounds and meanwhile the rank of the learned metric matrix is minimized.Experimental results on UCI datasets and two face databases demonstrate the effectiveness of the proposed method.Thirdly.the existing metric learning methods often rely upon too many features.and thus the performance becomes worse because of the non-informative features.In this thesis,we propose a sparse similarity metric learning method,which uses the sparse regularization to avoid relying upon too many features,and meanwhile applies posi-tive semi-definite constraint to smooth the learned similarity matrix.Since the fusion of different features can provide richer face description information.We further ex-tend the proposed algorithm into a multi-view metric learning method.Experimental results on the two face databases and two kinship face databases demonstrate the effectiveness of the proposed method.