Unconstrained Handwritten Numeral Recognition

In this thesis we mainly discuss general methods for unconstrained offline handwritten numeral character recognition. Offline handwritten numeral character recognition is a classic pattern recognition problem, which not only has practical application in banking, accounting and other fields needing mass numeral input, but also is a good test problem for new methods in pattern recognition and neural network etc.Handwritten character recognition has been extensively studied for many years and a number of techniques have been proposed. However, handwritten character recognition is still a difficult task and the result is still far from practical, especially when the character image is blurred and/or in multiple fonts. This thesis is focused on noisy unconstrained handwritten numeral character recognition problem. The main contents of the thesis are as follows.a) In preprocess stage, based on analysis of many normalization algorithm, the author propose a novel non-linear normalization method, which is the combination of point-density and line-density method.b) The author uses multi-feature extraction method to solve recognition problem. First, proposing a novel feature extraction method—Stroke Run-length Direction Feature, and then combining the Kirsch Mask Feature to improve recognition performance.c) System integrate many kinds of classifier to reduce the error rate and improve the reliability of recognition, including two neural network classifier— Back-propagation network and Self-Organizing Feature Map (SOFM) network, and two Bayesian classifier based on posterior possibilities.Based on above research, the author have finished an unconstrained off-line hand-written numeral recognition system, and acquired good result. Most of the ideas and algorithms can be applied to other pattern recognition problems.