Researches Of Wavelet Based Neural Network And Its Application On Information Processing

Wavelet analysis and artificial neural networks are the main components of the new generation of intelligent information processing technology. The wavelet function is localized both in space and frequency domain which makes the wavelet transform can effectively extract information from different resolutions by translation and expansion operations on wavelets. So the wavelet transform overcomes difficulties that the Fourier transform cannot solve and has become one of the advanced techniques of nonlinear science. Artificial neural network is a theoretical model abstracted from human brain neural network. It can acquire knowledge from environment by its ability of self-learning and storage knowledge by the connection weights between the neurons. Wavelet neural network(i.e. Wavelet network) is a combination of wavelet analysis and neural network. On one hand, it can maintain the multi-input parallel processing capability, self-learning ability, nonlinear mapping and fault tolerance of neural network. On the other hand, it can use the wavelet analysis with strong mathematical foundation to high dimension problems conveniently. Based on the research of classical wavelet neural network model, we choose the feedforward wavelet neural network as the main research object. In this paper, the initialization method of wavelet network with high dimensional inputs and tensor product form hidden wavelet functions; the structure and optimization algorithm of fuzzy wavelet network that combines the fuzzy mechanism and the algorithms of radial wavelet neural network are studied and analyzed. And applied them in signal prediction, system identification, pattern classification and so on, which makes these types of wavelet networks have a higher practical value and significance in the application. The detailed works are as follows:Firstly, the wavelet neural network with tensor product form wavelet activation functions(WNN_M) is the earliest proposed model that combines wavelet analysis and artificial neural network, whose translating and scaling parameters as well as linear weights are all adjustable. This kind of model is more flexible in applications and makes approximation of nonlinear systems with a less number of wavelets becomes possible. Q.H. Zhang has proposed a heuristic initialization procedure(HIA) which has been used widely in applications by this model. In order to get more ideal initial parameters making the training process converge more quickly, a novel initialization approach(CIA) which bases on the time-frequency localization of hidden wavelet functions is proposed. The approach determines the network’s hidden neuron number using a clustering algorithm by the help of a pre-defined threshold vector that according to the statistical characteristics of multidimensional input data, and determines the initial value of corresponding translating and scaling parameters basing on the cluster means and radii in terms of the time domain window width of the hidden wavelet functions. In the experiments of three kinds of different time series predictions, the initial errors of WNN_M-CIA are obviously lower than that of WNN_M-HIA. After several iterations of gradient descent method, the accuracies of CIA are still better than that of the HIA, which demonstrates the validity and rationality of this initialization algorithm.Secondly, considering the trend of combining fuzzy theory with wavelet analysis and neural network, a fuzzy wavelet neural network model based on TSK fuzzy system and wavelet network is proposed. A wavelet network with continuous parameters and tensor product form multidimensional wavelet activation functions is employed as the conclusion part of the TSK fuzzy model which replaces the traditional linear function. After analyzing the difference between the proposed model and the existing fuzzy wavelet models, a hybrid learning algorithm integrating the particle swarm optimization and online gradient descent algorithm is employed to training the proposed model, and applied it to nonlinear system identification. Simulation results of two identification examples illustrate that compared with existing models the proposed FWNN and algorithm can make a better identification performance even with fewer fuzzy rules and parameter numbers.Thirdly, based on Kohonen self-organizing map(SOM) neural network with competition mechanism, a novel approach for license plate slant correction and character segmentation is proposed. In this method, the pixels of characters are clustered into seven categories according to Euclidean distance of the coordinates and the slant plate can be corrected according to the tilt angle obtained from the neuron weight vectors. In addition, the weight vector obtained by SOM algorithm can be further used to segment the license plate characters by using the shortest distance method. The effectiveness of the proposed algorithm is illustrated by examples.Lastly, a novel self-creating disk-cell-splitting algorithm of radial wavelet neural network is proposed for license plate character recognition. Basing on the competitive learning of SOM, the proposed algorithm can bore neurons(i.e. the hidden neurons of RWNN) on a disk adaptively according to the distribution of input data and learning goals. The “circle neighbor strategy” employed to initialize weights of new neurons can make an ordered topology on the disk as well as utilize the trained weights which ensures an efficient training process. As a result, a disk map of input data and a RWNN model with proper architecture and parameters could be determined for the recognition task. Experiments of recognition of English letters and recognition of numbers or English letters are implemented for test the effectiveness of the proposed learning algorithm. Compared to the classical radial basis function(RBF) network with K-means clustering and LS method, the proposed algorithm shows better performance in the recognition problems even though fewer hidden neurons are employed.