Number Image Recognition Based on Neural Network Ensemble

Handwritten number is hard to recognize due to there existing much noise, and the shape, size, thickness and position of each number may be different. This paper investigates the recognition of number image of fixed pixels based on the neural network ensemble. Firstly, the Bagging technique is used to obtain the training samples, secondly, the discrete Hopfield neural network is employed to remove the noise among the samples and associate the samples, then 20 single three layers feed forward neural networks are built up to construct neural network ensemble to obtain the recognition results through voting. The research shows that the fault-tolerant and generalization ability of the neural network ensemble is superior to the single best model for visual number recognition.