| This dissertation is a study on algorithms and their applications of multilayer feedforward neural networks with matrix inputs.As we know,the input of traditional artificial neural network is vector form,and the image is usually represented by matrix.Therefore,in the process of image processing,the image will feed into the neural network in vector form,which will destroy the image structure information,and thus affect the effect of image processing.Moreover,the input of 2D-BP network is single channel input,and can not directly process color images.The image needs to be grayscaled before classification.Based on the above problems and analysis,the multi-multilayer feedforward neural network algorithms are studied and discussed in this dissertation.Also this dissertation addresses the applications on image classification using the algorithms.The main contents are as follows:1.We propose a multi-layer 2D-BP network algorithm based on the single hidden layer 2D-BP network,aimed at improving the generalization ability of the single hidden layer 2D-BP network and improving the ability of the network to extract the image features.Because the 2D-BP network is single hidden layer forward neural network,the generalization ability of the network is often insufficient.We extend the depth of the network and propose a multi-layer 2D-BP neural network algorithm.The experimental results show that the multi-layer 2D-BP neural network algorithms have better feature extraction ability compared with the single hidden layer 2DBP network image classification algorithm.2.The input of 2D-BP network is single channel,which cannot directly process color images.In order to overcome this difficulty,we propose a new three channel 2D-BP network,which extend the width of the 2D-BP network.For color image classification problem,the three channel 2D-BP network image classification algorithm proposed in this paper can extract the features of the images better and achieve better experimental results. |
PDF Full Text Request