Deep Learning based Handwritten Devanagari Character Recognition using Raspberry Pi 3

Abstract

A real time handwritten Devanagari character recognition framework using Rasp- berry Pi 3 is proposed. It has a Graphical User Interface application which takes image of character and trained deep convolutional neural network model as input and displays the output of character. The baseline deep convolutional neural net- work model is modified by changing its width i.e. number of filters and optimization function while keeping other hyper parameters like filter size, number of layers etc. constant. This is also helpful in understanding the impact of width of model on its accuracy that more width is not always better as training time increases with increase in width and eventually, accuracy starts decreasing. Large number of trainable parameters with increased width also make the model susceptible to overfitting which is resolved by dropout technique. The highest accuracy achieved is 98.75%. By loading this model, our application recognizes the character. Machine simulation and automatic recognition of Devanagari characters is useful in areas where the data present on paper has to be transferred to machine- readable format. It has a variety of practical and commercial applications in post offices, banks, libraries and publishing houses. Traditional methodologies used for Devanagari character recognition requires the extraction of statistical features and structural features of character. The accuracy of such classification was majorly dependent on feature descriptors. However, the nonlinear information processing used by deep learning models for feature extraction has been able to overcome these challenges. Among them, Convolutional Neural Networks (CNN) have be- come the leading architecture for image recognition and classification tasks.