Research On Feature Extraction And Metric Learning Algorithms In Face Recognition

Face recognition is a very important perception ability of human. With the development of computer technology, the human desire to give the computer the ability of face recognition, which gives rise to the research of face recognition technology. The research of face recognition has great theoretical meaning and application value. It can not only promote the psychology study of human face recognition mechanism, but also promote the study of image recognition and the realization of artificial intelligence. Face recognition technolory has been widely applied in civil security, public safety, entertainment, customer service and other areas. However,. Therefore, it is a very valuable research topic.Facial images will be influenced by illumination, pose facial expression, age and other external or internal factors. Moreover, the faces of some people are very similar by itself. These factors will increase the similarity of facial images of different persons and the differences of facial images of the same person, which causes a great deal of difficulty for face recognition. Therefore, face recognition is very challenging. Face recognition technology has not reached the level of demand in many applications, and thus need continuous innovative research.This paper focuses on feature extraction and recognition, which are the core of face recognition system. For feature extraction, this paper has studied the Gabor-based features and LBP-like features which are the main features in face recognition. This paper has analyzed their advantages and problems, and then proposed the improved features. For recognition, this paper has studied the metric learning methods, and proposed a probabilistic metric learning framework and two metric learning algorithms under this framework. The research work of this paper is listed as follows.(1)To solve the problems of slow extraction speed and high feature dimension which Gabor-based features suffer from, this paper proposed the HGMP (Histogram of Gabor Magnitude Patterns) feature. HGMP takes the Gabor filter bank as the code dicitionary for encoding the image points, and takes the filtering as encoding. To enhance the robustness and discrimination of the codes, HGMP further applied the orientation normalization and scale non-maximum suppression on the filter responses, which are very simple. The experimental results on the FERET and LFW databases validate that HGMP achieves good recognition performance while accelerates the extraction speed and reduces the feature dimension.(2) Based on HGMP, this paper introduced the log-Gabor filter bank to replace the Gabor filter bank and proposed the HLGMP (Histogram of Log-Gabor Magnitude Patterns) feature. Relative to the Gabor filter bank, the log-Gabor filter bank will retains more texture details and are more flexible to design, and thus it is more suitable to serve as the code dictionary. The experimental results show that HLGMP outperforms HGMP, and it outperforms all other competitive features on the LFW database.(3) This papar discovered that LBP-like features mainly adopt difference descriptor, which is correlative and causes the codes distributed very non-uniformly. Non-uniform distribution will reduce the discrimination of features. To solve this problem, this paper proposed the PLBP (Principal Local Binary Patterns) feature. PLBP adopts PCA (Principal Component Analysis) transform to the difference descriptor, which reduces the correlation and makes the codes distributed uniformly. Moreover, PLBP can utilize much larger neighborhood for difference discriptor which contains more information, and it can enhance the robustness of codes. The experimental results on the FERET and LFW databases validate that the codes of PLBP are distributed more uniformly and are more roubst to illumination variation, and thus PLBP achieves very good recognition performance.(4) This paper discovered that LBP-like features all adopted hard-coding, which will lose information and is sensitive to noise and illumination variation. To solve this problem, this paper proposed a probabilistic soft-coding method. It introduced the code confidence, which provided more information, and enhanced the influence of important image points, while reduced the influrnce of unstable image points. Therefore, the soft-coding will enhance the robustness and discrimination of the codes. The experimental results show that soft-coding can improve the performance realtive to hard-coding, especially when there is illumination variation, and the soft-coding PLBP achieves the best performance for feature matching on the LFW database.(5) This paper analyzed the metric learning problem from the perspective of probability and proposed a probabilistic metric learning framework. The framework relates metric to a hypothesis parametric probability distribution and constructs the optimization problem through the creterion of maximum likelihood or maximum a posteriori probability in paramter estimation. Under the probabilistic metric learning framework, this paper proposed a probability hypothesis of exponential function form, and deduced a simple EDML (Exponential Discriminatant Metric Learning) algorithm and LiDML (Linear Discriminatant Metric Learning) algorithm. This paper also introduced a priori probability hypothesis, and proposed the regularized EDML (REDML) and regularized LiDML (RLDML) algorithms. Regularization can suppress the influence of noise information in the training data and improve the performances of algorithms. The experimental results on the LFW database validate that EDML outperforms LiDML and regularization improves the performance. REDML achieves very good performance. Incoporating the HLGMP and soft-coding PLBP features proposed in this paper and LBP feature, REDML achieves the state-of-the-art performance on the LFW database.