As a kind of non-contact biometric technology, face recognition has a wide range of applications in military, economic, public security and other fields. At present, face recognition has become a hot research topic in pattern recognition, computer vision, im-age processing, neural network and so on. Because of the influence of illumination, facial expression and other factors, face recognition has become a very complex and challeng-ing problem. In order to reduce or eliminate the influence of complex illumination and to improve the recognition rate of face recognition, this paper try to improve the face recog-nition system from three main steps including:pre-processing step, feature extraction step and classification step. Meanwhile, since human beings are creating more and more data, the training samples of face recognition are greatly increased. This paper also discusses the optimization algorithm for large data. The main research contents of this paper are summarized as below.1. The illumination invariant extraction based on the edge-weakening image filter The aim of pre-processing is to remove the illumination conditions and obtain the original features of the human face, which is the illumination invariant of the face image. Quotient image method is an illumination invariant extraction method based on the Lambertian lighting model, and the algorithm can significantly improve the rate of face recognition under complex illumination conditions, and has very low computational complexity. However, the method still has some disadvantages, such as amplifying the high frequency noise in the low SNR region, and hard to keep edge information well. In this thesis, a self adaptive quotient image algorithm based on edge weakening filter is proposed, which can reduce the high frequency noise, and can better preserve the edge information such as the area in eyes, nose, mouth and so on, it can effectively strengthen the original feature, so as to improve the final recognition rate.2. local feature extraction method based on self-learning Good face representation is the key factor for efficient face recognition. The local feature extraction algorithm is a kind of main feature extraction algorithm. Local feature extraction method describes local change of pixel value. It is a very simple and effective method, but the local pattern of pixels in proposed by human experi-ences. In this paper, a novel method of local feature extraction is proposed, which is based on the method of self-learning.The proposed method drive the local fea-ture pattern by training instead of human experiences.By using this method, the difference between images from the same person is further reduced, and the differ-ence between images from different people are enlarged. Through this method, the recognition rate is highly improved and more robust to changes of illumination and expression.3. feature extraction and classification identification method based on the unified the-ory Classification and feature extraction is a relatively independent two modules in face recognition system. In many studies, the two steps are optimized seperately. How-ever, as part of the face recognition system, these two parts are closely related. The two complement each other, the good feature can enhance the classification per-formance, the good classifier can reduce the requirement of feature extraction. In this paper, the internal principle of these two modules is analyzed in detail, and the unified principle is proposed, that is, the distance of the point to the subspace. The method of feature extraction and classification based on the same inner principle is proposed. The experimental results show that the extraction and classification method based on the same principle can improve the performance of face recogni-tion.4. the mini-batch quasi-Newton optimization method for large scale face recognition With the development of science and technology, the big data era is coming, the exponential increase in the amount of data brings a new challenge to existing al-gorithms. In this paper, a new mini-batch quasi-Newton optimization algorithm is proposed for large data training. The method try to use small batch of training samples to calculate the parameters, and to optimize the final classification model. The large amount of computation time can be effectively reduced by using a small number of samples while remains the final recognition rate. |