Research On Face Recognition Based On Improved FDA And RBF Neural Network

With the development of technology, information security is desperately demanded by all individuals worldwide. Face recognition becomes the most popular research field of biometrics for its directness, friendliness and ease of use. Combined with statistical theory, machine learning and information theory, many face recognition algorithms have been proposed, among which face recognition based on features is most widely used.Firstly, this paper gives a brief introduction to the main algorithm, key steps, performance indicators and image pre-processing methods in face recognition. Then, mainly focus on performance comparison of face recognition based on overall features and local features.Feature extraction based on subspace is recognized as an effective method to extract the overall features. In this paper, comparisons are conducted in aspects of algorithm theory, ways of application in face recognition and simulation performance among Principal Component Analysis,2Dimensional Principal Component Analysis, Fisher Discriminate Analysis, Singular Value Decomposition and Independent Component Analysis. Simulation results indicate that algorithms mentioned above are excellent at feature extraction and data compression.Multi-channel Gabor filters is widely used in local features extraction for its effectiveness in simulating the recognition mechanism of human eyes. Firstly, the principle of Gabor transform is researched. And then, Gabor filters with a scale of5and frequency of8are design to extract features. However, Gabor features are too many to process for MATLAB, so down-sampling and Linear Discriminate Analysis are used to compress the data. The simulation results show that the multi-channel Gabor transform can achieve higher recognition rate, but too time consuming.Feature classification is as important as feature extraction in face recognition.3-order nearest classifier is used when doing the comparison in feature extraction algorithms. Distance classifiers are traditional and simple ways for classification. In the last part of this paper, Bayesian classifier、 RBF neural network classifier and Support Vector Machine classifiers are used to classify PCA features. Simulation experiments show that RBF neural network is superior to other classification method.