The Study Of BP Learning Algorithm Of Artificiai Neural Network And Its Application In Face Recognition

Artificial Neural Network (ANN) is a model which has nonlinear processing capability obtained by the simulation of biological nervous system. It is connected by a large number of neuron cell models according by a certain rules. It gets complex nonlinear calculation ability by mapping a large number of simple calculations of the neuron model. Therefore, it is widely used in many fields, like pattern recognition, image processing, data mining, signal processing, and automatic control and so on. One of the most commonly used artificial neural network model is the Multi-Layer perceptron network model (MLP). It is composed by Multi-layer neuron cell models, and each neuron cell model uses the Sigmoid activation function. The learning algorithm of MLP network model commonly is back-propagation algorithm (BP), so it is known as BP neural network. The Traditional learning algorithm of BP is gradient descent algorithm, which the learning factor is a fixed value set by experimenter. But because of the high-dimensional complexity of the error function and the learning factor set incorrectly, in the learning process of gradient descent algorithm has many shortage like the error function convergence too slow and even does not converge, has shock, and easy to fall into local minimum points and so on.Based on the three layer fully connected MLP network model, aimed to the shortage of traditional BP gradient descent algorithm, this article propose several improved algorithm, including the steepest gradient descent algorithm, conjugate gradient algorithm, output weights optimization algorithm, optimal learning factor algorithm, multiple optimal learning factor algorithm.Steepest gradient descent algorithm calculates the learning factor by the gradient of error function to learning factor to be zero; conjugate gradient algorithm calculates the learning factor by using the conjugate optimization in mathematical, it use the gradient information calculate the conjugate gradient and then replace the original gradient; output weights optimization algorithm calculates the output weights by constructing linear equations, because it only optimize the output weights, so it can be used combine with steepest gradient descent algorithm or conjugate gradient algorithm; optimal learning factor algorithm calculate the learning factor by higher derivatives of error function gradient; multiple optimal learning factor algorithm extend the learning factor to be a vector, each element correspond to a learning factor of hidden layer nodes. The experimental results in the public datasets show that the improved algorithm can effectively solve the shortage of traditional BP gradient descent algorithm.Neural network has been widely used in face recognition field. This paper designs a classifier based on MLP neural network model which used BP learning algorithm, and use in the face recognition systems under complex lighting conditions. This paper proposed a face image feature extraction method. In this method, first, face images are filtered by multi-directional, multi-scale Gabor wavelet filters; then, the filtered images is binary by adaptive threshold value; at last extract the local information as the feature vector of the face images. In order to reduce the computational complexity of the neural network classifier, feature vectors were reduced dimensions by using principal component analysis (PCA), and then use the neural network classifier learning and classification. This method combined with the advantages of Gabor wavelet filters; adaptive threshold binarization and neural network classifier, the experimental results on Yale extend B face dataset show that the method can achieve good classification and recognition results.