Research And Implementation Of Feature Extraction Algorithm In Face Recognition

Face recognition is a biometric identification technology with great developmental potential. And face recognition has broad application prospects in many fields, such as bank, police system, social welfare safeguard, etc. Over the past several decades, great progress and developments have been made in face recognition. At present, face recognition achieves satisfactory performance under the condition of controlled and cooperative. However, it remains a challenging project due to other complex factors, including illuminations and facial expressions, which lead to a sharp decline in accuracy. Therefore, it is quite significant to develop a feature extraction algorithm with a high robustness and a better characterization capability. In this thesis, we focus on the feature extraction method which can apply to complex illumination, and implement it in a prototype system of face recognition. The main works are as follows:(1)Some related works of face recognition, such as the advantages in biometrics recognition technology, the application fields and the current problems, are analyzed and summarized. And the domestic and international research status of feature extraction algorithm are introduced. Meanwhile, the feature extraction and feature selection methods are analyzed..(2)The majority of existing feature extraction algorithms are easily influenced by the external factors, especially illumination, which decreases the accuracy of face recognition. Histogram of oriented gradient(HOG) can redu ce the interference caused by the illumination, because of strong illumination robustness. However, traditional HOG only considers the impact of the four pixels situated in horizontal and vertical direction when obtaining the gradient magnitude. The stability of algorithm can not be guaranteed when the external environment changes. Thus, an improved HOG feature extraction algorithm based on Haar characteristics is proposed. Compared with traditional HOG, When calculating the gradient direction and amplitude, the proposed algorithm considers the influence of 8 pixels. Meanwhile, because of the simple and fast operating of Haar-like features, the proposed algorithm inducts Haar into HOG. Then the proposed algorithm shows four groups of Haar feature encoding models, and extract the face features according to the improved HOG. Experimental study shows that the method has better illumination robustness and the accuracy of face recognition is improved in complex environment.(3)For the problem of large feature dimension and redundancy feature. This thesis uses Restricted Boltzmann Machine which has strong representation capability and reasoning capability as a solution. However, RBM may fail to model the data distribution in training process when the parameters are unsuitable for data sets or the structure of RBM. Consequently, the thesis proposes a feature selection algorithm based on rough set and restricted boltzmann machine. First of all, rough set theory is used to analyze the relationship between samples; as a result, the concise decision rules and the attribute membership are extracted from the given training data. Afterwards, the initial parameters of RBM are determined by the concise decision rules and the attribute member ship. Meanwhile, in order to optimize the RBM algorithm, a new weight updating rule is presented, which improves the weight by the initial weights and the reconstruction error. Experimental results show that the proposed algorithm can reduce the iterations of RBM training and the reconstruction error, as well as extract the better features with low dimension and better separability.(4)Finally, a prototype system of face recognition is designed and implemented. And the proposed feature extraction algorithm and feature selection algorithm can be implemented in the system. Meanwhile, in order to improve the face recognition rate, the system uses deep belief network(DBN) to select the feature. Lastly, the system is tested in function-level and system-level, and the tested results are analyzed.