摘要
采用模糊统计方法获得有限对离散的隶属函数关系，并用神经网络将此关系综合成连续函数表达形式。这一方法进一步完善了隶属函数的统计求法，丰富了经典插值方法，有利于模糊数学理论应用于希望实现并行处理的模式识别等研究领域。我们将这种方法应用于语音识别研究，取得了较好的结果。
In this paper, a new approach to implement the membership functions with functional-link networks hased on fuzzy statistical method is proposed. Functional-link network, proposed by Dr. Sobajic in1988, is a kind of plane network. It posseses a strong capability of interpolation and mapping. Patterns withindistinct features could be described and classified in higher dimensional vector space with this kind of network because of its functional extending function.Although membership function is the key to introduce the fuzzy set theory to pattern recognition system,it is difficult to set up membership functions for fuzzy events with a large number of numerical data sampledfrom physical transducers. Fuzzy statistical method is very effective in the task, but the membership functionsimplemented with it are discrete. In many pattern recognition systems, continuous instead of discrete membership functions are wanted. Baed on the capability of intetpolation of functional-link network, discretemembership functions derived from fuzzy statistics are used as training samples. When the network converges, not only the continuous membership functions implemented with the network, but also a kind of parallel processing structure is obtained. This kind of structure is very significant to parallel pattern recognitionsystems.Membership functions based on functional-link networks.have been used in speech recognition systemusing line spectrum pair parameters as speech features.The speech recognition results are comparable to those obtained from Hidden Markov Model system,a famous speech recognition method,in which cepstral parameters are used as speech features.IN text-independent speaker identification system which consists of 42speakers'templates, the performance of the system is better than that of original one in which the membership functions were supposed to be normal distributions.Apart from the effect on establishing membership functions, the method of functional interpolation using neural network in also a novel approach besides classical interpolations.Of course it could be used to solve the problem of functional interpolation.
出处
《南京大学学报（自然科学版）》
CSCD
1996年第3期421-426,共6页
Journal of Nanjing University（Natural Science）
基金
国家自然科学基金
关键词
模糊统计
隶属函数
神经网络
模式识别
Fuzzy statistics,Membership function,Neural network,Pattern recognition
