International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017
p-ISSN: 2395-0072
www.irjet.net
Handwritten Digit Recognition Using Convolutional Neural Networks T SIVA AJAY1 School of Computer Science and Engineering VIT University Vellore, TamilNadu,India ---------------------------------------------------------------------***--------------------------------------------------------------------Abstract - Handwritten digit recognition is gaining a huge demand in the branch of computer vision. We are going to implement a better and accurate approach to perceive and foresee manually written digits from 0 to 9. A class of multilayer sustain forward system called Convolutional network is taken into consideration. A Convolutional network has a benefit over other Artificial Neural networks in extracting and utilizing the features data, enhancing the knowledge of 2D shapes with higher degree of accuracy and unvarying to translation,scaling and other distortions. The LeNet engineering was initially presented by LeCun et al in their paper.The creators excecution of LeNet was primarily focused on digit and character recognition.LeNet engineering is clear and simple making it easy for implementation of CNN’s. We are going to take the MNIST dataset for training and recognition. The primary aim of this dataset is to classify the handwritten digits 0-9 . We have a total of 70,000 images for training and testing. Each digit is represented as a 28 by 28 grey scale pixel intensities for better results. The digits are passed into input layers of LeNet and then into the hidden layers which contain two sets of convolutional,activation and pooling layers. Then finally it is mapped onto the fully connected layer and given a softmax classifier to classify the digits.We are going to implement this network using keras deep learning inbuilt python library.
decision than a perceptron in the first layer. The layers away from the first layer make progressively more complex decisions compared to the first layer. For learning purpose we should continuously change the weights so that the network finds out the aggregate and compares it with a threshold value of bias. If a small change in the weights modifies the output in the direction we want to proceed then we can use small weights or we can take large weights for training ,this method is like hit and trial which we use in solving higher degree polynomials. The architecture of neural networks is divided into three categories,the input layer neurons, the output layer neurons , the layers in between input and output layer called as hidden layers. Sometimes the networks have multiple layers they are coined as Multilayer perceptrons or MLP’s. The input layer of our network consists of input neurons encoding the values taken from input pixels of our handwritten digit.Our training data which is fetched from MNIST data set consists of many 28 by 28 pixel images and so input layer contains 784 input neurons. The second layer of our network will be the hidden layer and it takes the aggregated output of first layer and applies activation function to detect the pattern of input images. We will experiment with different values for the number of neurons in the hidden layer. Next coming to output layer of our network contains 10 neurons, each neuron if fired gives any output between 0 to 9 ,so this is the basic working of neural networks for an image. Similar to those of artificial neural networks but with a small modification comes Deep learning networks also termed as Convolutional networks with a higher degree of accuracy for classification. In Deep networks features are extracted not from the entire domain of input but several features are extracted from parts of the domain. Our deep networks use LeNet architecture which enables the networks fast to train and obtaining better results of classification. Keras is a highlevel deep networks library written in python and capable of running by using backend as either Tensor Flow or Theano. It was developed for focusing mainly on deriving fast training. Supports both convolutional and recurrent networks and combination of both also. The main principle of this library is modularity , which is understood as a sequence of fully configurable modules that can be combined together with little constraints on them. In particular neural layers, cost functions, optimizers, activation functions are all modules that you can combine to form a new network. The main aim of keras is modularity ,a way of architecting the
KeyWords:ConvolutionalNeuralNetworks (CNN’s),LeNet, Artificial Neural networks 1.INTRODUCTION In our half of the globe of our mind otherwise called V1 ,contains millions of neutrons with billions of connections between them. The thought is to take digits into consideration and build up a framework which can gain from these. At the end the neural system utilizes the cases to naturally construct rules for perceiving handwritten digits. There are two types of neurons accessible in our brain are perceptron, sigmoid neuron. To calculate the yield we will present weights computing the significance of the separate contributions to the yield. The neuron’s give an output of 0 or 1 if the weighted sum is below or above some threshold value. Various decision making models are formed by different weights and threshold values. In the network the first layer of perceptron’s that makes very simple decisions ,by multiplying the weights with the inputs. In this way a perceptron in the second layer can make even more complex
© 2017, IRJET
|
Impact Factor value: 5.181
|
ISO 9001:2008 Certified Journal
| Page 2971