Face Feature Extraction And Recognition Based On Ensemble Learning

Face recognition is one of the intensive research topics of pattern recognition, computer vision and human-computer interaction. With the focus on security issues, application of face recognition technology in public security, information security and business security is very broad.The complete process of face recognition can be divided into three stages:face detection and location, image preprocessing, face feature extraction, face recognition. Among them, face feature extraction plays a very important role in face recognition. Subspace analysis method is currently the most widely used a feature extraction method, which has been widely research and development. At the same time, a single learning algorithm has poor generalization ability, so its performance cannot meet the requirements. Ensemble learning is an important research direction in machine learning research, use multiple learning to learn, and can effectively improve the learning performance and generalization ability of the algorithm.This paper first introduces the research background, basic theory and method of ensemble learning. It made effective explorations on a combination of feature extraction and ensemble learning in face recognition. The main work of the thesis includes:1. Based Bagging LDA ensemble and its application in face image recognitionLinear discriminant analysis method is a global face feature extraction method. When LDA deals with the high dimensional face image, recognition instability of the local change is a challenging problem. In order to address it, a new method named Bagging_LDA is proposed based LDA and ensemble learning method Bagging in this paper. Different from the traditional Bootstrap Aggregating(Bagging) which completely randomly samples from the whole face image. Bagging_LDA random sampling on each local region partitioned from the original face image. More specifically, a face image is divided into several sub-images set, and then constructs a training set of LDA on Bootstrap Sample set and initial set of sub-images. In the stage of classifiers ensemble, a component classifier with best performance is selected from a set classifiers. Using the method of weighted voting to combine the results of classifiers. Experiments on ORL and YALE face database show that the proposed bagging_LDA method is effect in recognition performance.2. Discriminant analysis method ensemble Based on RSMUnsupervised Discriminant Projection is a kind of dimension reduction method preserving local and non local information of. For unsupervised discriminant projection (UDP) algorithm classification ability is weak and unstable performance problems, a based RSM discrimination analysis ensemble method(DAC_EL) is proposed. At the training stage, based on random sampling feature of training set, lots of UDP projecting transformations are gotten from RSM (Random Subspace Method),that is, gotten many differences and complementary projection subspace. By test samples, the accuracy of individual classifiers in the ensemble are gotten, which are used as their weights in ensemble step. The experimental results on ORL and YALE illustrate that the performance of DAC_EL method is superior to single UDP classifier and majority votes UDP ensemble, and significantly improve the classification ability and stability of the face recognition.3. Linear discriminant analysis and ensemble Based on Prior distributionLinear discriminant analysis method is a classical linear feature extraction method in the 3. Linear discriminant analysis and ensemble Based on Prior distributionLinear discriminant analysis method is a classical linear feature extraction method in the field of pattern recognition, obtaining the optimal projection transformation W by maximizing the between-class scatter and minimizing class scatter. In practical applications, the sample data are easy to obtain, and the probability of sample data are not easy to obtain. In the paper, using the method of Boosting, a linear discriminant analysis ensemble approach based on prior distribution is proposed. The method obtains the probability of the samples that are hard to process by the previous classifier through Boosting. The probability of the pattern sample is applied to the LDA feature extraction and the process of constructing minimum distance classifier. In each iteration, the training sample weighting error rate are calculated as the weight of the base classifier. And update the weights of the training sample as probabilities of sample appear of the next iteration. The minimum distance classifier uses two output forms to get two kinds ensemble method BP-LDA-1method and BP-LDA-2methods. The experimental results on ORL and FERET illustrate that the performance of BP-LDA method is superior to PCA and LDA. Experiments show that the algorithm can effectively improve the classification performance of face recognition.