Devanagari Character Recognition Using Artificial Neural Network
Vasu Negi, Suman Mann, Vivek Chauhan · International Journal of Engineering and Technology · 2017
Devanagari is one of the Ancient Scripts that is in regular use in India as well as Nepal.Devanagari is currently used by more than 120 languages including Hindi, Nepali, Marathi, and Pali which defines the prevalence and the domination of Devanagari [1].Immense popularity of this script must be taken care by the advance technologies of our present world so that we are able to connect to the real world to a greater depth.Artificial Neural Network is one the fastest growing technologies that has attracted a lot of attention in the field of Machine Learning.Various new technologies have developed to implement fast neural networks with little in depth knowledge requirements.We will be using Keras along with Theano which are python libraries for building our neural network.In this paper we constructed a simple artificial neural network using keras to recognize isolated Devanagari characters and also analyze the impact of variations in parameters in learning phase.Keyword-Devanagari, Character Recognition, Keras, Theano, Python I. INTRODUCTION Devanagari script is one of the oldest script that is regular use in India as well as some other south Asian countries.Devanagari script is used in languages and dialects including many major languages such as Hindi, Marathi and Nepali [1]. The world is currently moving towards digitization and thus we require means to translate documents written in Devanagari into a digital format so that they can be preserved, manipulated and shared seamlessly.Devanagari can be found to be used intensely in various areas of India from handwritten applications for installing a water connection or may be a letter of request to any concerned authorities, Devanagari is used by millions of Indians every day.Millions of documents are created written in Devanagari script every day, these handwritten documents must be taken care.Hence, we require efficient and flexible ways for doing so.Artificial Neural Network is one of the technique that can be applied to efficiently recognize the Devanagari characters.Artificial Neural Network is simply a network of interconnected nodes that provides classification and regression abilities to the machine.II. ARTIFICIAL NEURAL NETWORK Artificial neural network are simply a network of interconnected nodes that provides classification and regression capabilities to the machine.These nodes are supposed to mimic biological neurons in a normal brain.Same as biological neurons, nodes does not have any computation on its own rather they act as a group of linear functions.Each node has the task to throw an output when a certain level of threshold has been received from other nodes [3].The behaviour of each node can be defined by activation function which specifies.Activation function takes several inputs that comes from other nodes that are connected to a node and produces an output if a threshold is reached.A single node in itself cannot possess any computation capability.An ANN comprise of several nodes that are arranged in layers where each layer can comprise of several nodes.Each node in a layer is connected to another node in the adjacent layer.A simple neural network with 3 layers comprises of an input layer, hidden layer and an output layer.The input layer does the task of collecting the input and the output layer presents the result in the form of classification using one of the several output nodes.Each node in the output layer will act as an independent class which will classify the data into one of the classes.Hidden layer is the layer where the learning takes place.Each connection between nodes are assigned 'strengths/weights', usually assigned in a random fashion initially.These strengths attributed to each connection forms the basis of learning.These strengths are then incremented or decremented in way that the network is able to tune the output to a certain degree.A weight can also become zero denoting that the connection does not hold any importance.We will be using Keras and theano which are python libraries for building our neural network.These libraries are discussed in depth in later sections.