A Study of Character Recognition Using Neural Networks

To improve the recognition accuracy of the system with the identified the feed forward NN as classifier, the best possible feature set is to be obtained. Among the different feature extraction techniques, the statistics based zonal feature extraction technique is chosen for investigation. In zonal approach, horizontal and vertical feature extraction techniques are studied; both these methods have symmetry about x-axis and y-axis respectively. A new feature extraction technique, called diagonal feature extraction method is proposed in this thesis, which is symmetric about its diagonal. To further enhance the recognition accuracy of the hybrid feature based CRS without increasing the complexity of the neural architecture, a new training strategy, called novel training is proposed. To implement this novel training strategy, the alphabets with low recognition accuracy are identified. Low accuracy is defined using an adaptive threshold parameter. The characters having low recognition accuracy and their corresponding confused characters are identified. Generally, equal number of samples of all characters is taken for training. In this proposed novel method, more numbers of samples of the characters with low recognition accuracy and their confused characters are included for training. The optimal value for additional data sets, Δx of low accuracy characters is determined through an analysis. With the novel training strategy, the proposed CRSs – built one for uppercase and one for lowercase, are shown to exhibit improved recognition accuracy.