Research On Face Recognition Method Based On Metric Learning

Face recognition is an important biometric technology. It has been widelyapplied into every identification field, such as human-computer interaction, videosurveillance and electronic passports, for its easy capture method and convenientoperation. The variance of age, pose, illumination, cover, accessories and expressionin face image makes face recognition still a challenging problem in computer visionfield.The goal of metric learning is to learn a proper metric to measure the distanceof two instances. Metric learning is equivant to linear dimensionality reductiontechnology. In face recognition, the samples have high dimensions and usuallyinclude information which is irrelevant to face recognition. These informations canaffect the recogniton performance. By linear dimensional reduction technology ormetric learning method, we can exclude the irrelevant information and redundantinformation, and reserve the effective feature for face recognition. Thus we canimprove the recognition performance.The traditional metric learning method would be problematic to face recogniton.It can lead to a series of problems, such as overfitting, unrobustness to image noise,low training speed, and so on. To address these shortcomings, this paper researchesabout metric learning method based face recognition, and proposes a series ofsolutions to those problems.(1) Gabor feature–based fast neighborhood component analysis: This methodis proposed based on fast neighborhood component analysis model. Since fastneighborhood component analysis model would be overfitting for high dimensionaldata, in order to alleviate the overfitting problem, we use PCA to reduce thedimension of the data, and introduce a Frobenius regularizer into fast neighborh oodcomponent analysis model. To improve the performance of face recognition, we useGabor filter to extract the features from face images.(2) Fast neighborhood component analysis with spatially smooth regularizer:Since fast neighborhood component analysis is very sensitive to image noise, andspatially smooth regularizer is successful in image denoising, so we introduce thespatially smooth regularizer into fast neighborhood component analysis model toimprove its robustness to image noise.(3) Doublet and triplet based support vector machine: This paper first proposesa metric learning framework based on kernel classification. This framework learns ametric via kernel classification method. It can unify many representative andstate-of-the-art metric learning methods. Furthermore, we can propose new metric learning methods based on this framework. This paper proposes doublet and tripletbased support vector machine model with this framework. These two models can beeasily solved by many existing SVM toolboxes and solvers, and much faster thanstate-of-the-art methods in terms of training time.