Fuzzy Filtered Neural Network Approach towards Handwritten Numeral Recognition

This paper presents a fuzzy filtered neural network approach as an application to handwritten numerical representation. A multilayer feedforward adaptive network is used for training the model and for application of the fuzzy filters. Fuzzy filters are integrated with the neural nets for processing the physical data of the images available for the handwritten digits. The use of the fuzzy filters reduces the noise and redundancy present in the data which ultimately increases the performance of the model. This also helps in avoiding the high complexity of the neural network architecture which would otherwise be required for the same physical data. Different varieties of the fuzzy filters are integrated with the neural network separately and their performance is compared. The fuzzy filters with higher dimensionality improve the model recognition rate. One dimensional and two dimensional fuzzy filters are discussed and their performance is evaluated. Finally, genetic algorithm based fuzzy filtered neural networks are discussed for the application of recognition. They provide the highest recognition rate for the application.