Recognition of online Handwritten Bangla Characters using Hierarchical System with Denoising Autoencoders

This work describes the recognition of online handwritten Bengali characters using Deep Denoising Autoencoder with Multilayer Perceptron (MLP) trained through backpropagation algorithm [I]. Initial pre-training has been done to the Denoising Autoencoder with MLP trained through backpropagation algorithm, to bring the weights of the Deep network to some good solution and then pre-trained Denoising Autoencoders are stacked to form a Deep Denoising Autoencoder (DDA). A final classification layer makes DDA to a Deep Classifier (DC) followed by a final fine-tune that gives the best classifier for the job of classification of Bengali characters. The overall system is hierarchical in nature and the system has been trained in two phase where the first phase has trained a broad classifier and in the second phase class specific recognizer has been trained. At the testing phase in this hierarchical approach, first a broad classifier has been used to recognize broad classes like Vowel, Consonant, Special Symbol and Numeral for a novel test sample. Once the broad class gets recognized then a class specific recognizer has been used to recognize the exact character the test sample belongs. Recognition performance of the hierarchical system is 93.12%.