An efficient multiclassifier system based on convolutional neural network for offline handwritten Telugu character recognition

We combine the strengths of four different pattern analysis techniques to develop a powerful and efficient system for handwritten character recognition. The four techniques are: 1) Convolutional neural networks (CNN), 2) Principal Component Analysis (PCA), 3) Support vector machines, 4) Multiclassifier systems. The proposed system that embodies the above-mentioned four techniques is used for recognition of offline handwritten Telugu characters. Telugu aksharas of consonant-vowel (CV) type, with 36 consonant classes and 15 vowel modifier classes, are used for the study. Telugu dataset consisted of 47428 CV images in the training set and 5156 CV images in the test set. In addition to Telugu dataset, MNIST database consisting of 60000 digits for training and 10000 digits for testing was used in this study. The proposed system yields a performance of 98.5% on MNIST numeric data, 92.26% and 92% on consonants and vowel modifier of Telugu characters respectively.