The Research On Neural Network With Fractional Weight Function And Its Application In Texture Classification

After more than 50 years development, the neural network research has received giant achievements.However, it is difficult for traditional neural network learning algorithms (such as BP algorithm, RBF algorithm) to reflect the information of training samples, meanwhile,in practical applications, the traditional neural networks models are difficult to be settled. Based on these considerations, Daiyuan Zhang of Nanjing University of Posts and Telecommunications proposed a new algorithm-spline function weight neural network training algorithm. This training algorithm overcame the defects such as partial minimum frequently, low convergence speed, sensitive to initial value which exists in traditional neural network. It is proved in both theory and experimentation that the spline function weight neural network has better performance.Based on the research of Professor Daiyuan Zhang, this paper constructs a new kind of artificial neural network -rational spline function weight neural network which expands the spline function weight neural network theory. Using rational function as weight function,this paper constructs a new type artificial neural network-rational weight function neural network by means of reciprocal difference -continued fraction method, and then analyze the network error to combine with continued fractions function neural network. Finally, we can conclude that continued fractions weight function neural network has high accuracy and convergence speed through simulation experiments, which is compared with conventional BP, RBF neural networks algorithm in the mean square error and computational speed.Finally, the continued fractions function weight neural network's application in texture classification is given. Gabor wavelet transforms combining by Gabor filter and Wavelet Transform is used in analysis and processing of texture image. The construction of continued fraction function weight neural network texture classifier is given, Comparing with traditional ways, continued fraction function learning algorithm has a better classification accuracy and time efficiency.