A Lobe Component Analysis Based On-line Learning Algorithm And Its Application In Face Recognition

Among the issues in the research of pattern recognition,feature representation as well as extraction is one of the most important as the efficiency of the feature obtained is of critical importance to the final solution of the classification problem.In the field such as computer vision,data is prone to have extremely large dimensionality which requires the feature representation module should have the ability to represent the feature of the original data space with a low-dimension feature space.The subspace method is a collection of feature learning methods that have been extensively studied and widely used by researchers and it is proved to be efficient in solving the high-dimension pattern recognition problems.Traditional subspace methods apply batch calculation procedure to obtain the feature.Recently people have pay attention to the research of incremental subspace methods that have the ability to dynamically update the feature representation utilizing the newly obtain observation sample.So it is possible for the incremental subspace methods to be used in implementing on-line learning methods which can perform feature learning during the pattern classification procedure.The merit of on-line learning lies in that it can make full use of the input samples obtained when the pattern classification task is being performed.In some cases,the size of the available training samples is small or the samples are not representative enough.So,the features extracted from such sample set are not sufficient for designing a well-performed classifier.However,the on-line learning methods can overcome the insufficient training problem by its capability of on-line learning.Besids that,in nonstationary systems the probability distribution of the observation is changing such that the features eatracted based on the training samples are pron to be inappropriate for classifying the non-stationary input.In such non-stationary situation,the on-line learning methods are capable to catch the evolution of the observation by adaptively updating the feature model. Based on the Lobe Component Analysis(LCA),which is a newly published incremental method,we proposed an on-line learning method name LCA-BOLL(LCA Based On-Line Learning).Compared with other subspace on-line learning methods,the proposed one has several unique properties:(a) the feature component of LCA-BOLL has a class label,(b)LCA-BOLL combines both supervised learning and unsupervised learning,(c)in LCA-BOLL the two procedures of on-line feature learning and pattern classification are seamlessly combined.The proposed LCA-BOLL algorithm is applied in the problems of face recognition in static image and face recognition in image sequence separately.In the application of face recognition in static images,the LCA-BOLL algorithm is applied to solve the Small-Sample-Size problem which is encountered by learning methods when the sample size of training samples is too small.While in the application of face recognition in image sequences,we used LCA-BOLL to solve the problem of large variation of face pose in image sequence.Experiments in both two applications proved that LCA-BOLL is sufficient and effective in solving the problems we focused on.The efficiency of LCA-BOLL's on-line learning ability is verified by its performance in the application of face recognition.