Fast Face Recognition Research Via Patch Based Ensemble Learning And Collaborative Representation

Face recognition has been an active research topic in the field of pattern recognition and image processing for a long time. Currently, there are many algorithms that can achieve satisfactory recognition rate, however, most of these methods are used under strict limit of conditions, such as frontal face and good illumination condition. In addition, most of these methods have low efficiency, which cannot meet the real-time requirement. How to improve the recognition rate and balance the efficiency? it is a research problem which has important practical and theoretical value. This paper proposes a patch based ensemble learning framework for the problem of illumination variation, expression variation and small sample size in face recognition which can well balance the performance and efficiency.The framework mainly contains three components:(1) the proposed effective illumination insensitive face feature enhancement algorithm:"GDP-face", it based on rotation invariant local binary pattern and gradient direction which could well extract the face feature and remove most of the influence of illumination variation. (2) the proposed fast collaborative representation based classification (FCRC), it improves the traditional sparse representation based classification which could significantly decrease the computational cost while slightly improve performance. (3) patch based ensemble learning method, it first partitions the face image into several overlapping patches, then apply the proposed fast collaborative representation based classification on each patch, finally, using an expression insensitive two stage voting method to combine the classification result of each patch, which could further improve the recognition performance under varying expression.The proposed patch based ensemble learning framework improve the performance for face recognition under complicated environment from feature enhancement, classification and ensemble learning, respectively, meanwhile, the framework has high efficiency by using the proposed fast collaborative representation. Extended experiments show the effectiveness of the proposed patch based ensemble learning face recognition framework by comparing it with other patch based algorithms.