Learning Algorithms And Application On Image Classification Of Multilayers Feedforword Neural Networks

In this dissertation,we study the learning algorithm of multilayers feedforward neural networks and its application on image classification.Image classification is one of the basic tasks of computer vision,which uses computer technology to extract the various features from the image to classify them correctly.Based on the applications of deep learning and feedforward neural networks algorithms on image classification,we propose two algorithms,i.e.,learning algorithm for multilayers feedforward neural networks with matrix input and learning algorithm of neural networks with the weighted hybrid feature vectors.The main contents are as follows:1.We proposed the learning algorithms for multilayers feedforward neural networks with matrix input.Generally,there are two kinds of inputs for feedforward neural networks: one is the vector form,and the other is the matrix form.For image processing,if the number of networks layers is same,the learning ability of feedforward neural networks with matrix input is better than that of vector input.The reason is that the neural networks with matrix input overcomes the problem of destroying the spatial structure in images.Based on the fact that multilayers neural networks has better feature extraction ability and generalization ability,we extend the neural networks of single hidden layer with matrix input to the multi hidden layers and use the traditional back propagation algorithm to train our model.Through the comparision of several databases,the experiment results show that the proposed algorithm possesses good performance compared with some existing methods.2.We proposed an algorithm of neural networks with the weighted hybrid feature vectors.In the learning process of feedforward neural networks,the next layer of network feature vectors can be regarded as a special mapping of the previous layer.With the networks going deeper,the feature vectors extracted from networks will be more abstract.When mapping the output layer of neural network onto the basis of second reciprocal layers,it is easy to ignore the effect of other layers for networks tasks.Based on this,we map the two layers of feature vectors through a convex combination to the output layer.In other words,the output layer neurons in the networks are not only connected with neurons in the last but one layer,but also connected neurons in other layers.This convex combination network mapping can be regarded as a mixing process of low level feature vectors and high level feature vectors.This algorithm possesses good classification results through the experiment on the FERET face database.