Handwritten Character Recognition using Convolutional Neural Network
Sagar Kumar Singh, Rahim Alam, L. Sujihelen, Josila Grace L. K, Mercy Paul Selvan, S. Jancy
Abstract
The proposed work aims on the task of identifying handwritten alphabets using a predictive designed CNN model. The visual image of alphabet character can be stored digitally for future use. The collection of handwritten english alphabet data-set is available at Kaggle website. The data-set file includes twenty-six folders containing twenty-six english alphabet images of 28×28 pixel size. Every image is stored as a default color as greyscale. The data-set of images is classified into 20% of test data and the remaining 80% as training data for our neural network. This model is constructed with the help of TensorFlow as it's back-end. It contains Convolution filters with 3 layers of fully integrated visualization. The intake of this Convolutional Neural Network model are categorical values. So, greyscale images is converted to categorical values as vector form. Filter layers in CNN extracts the image elements from all the image data. After image integration for feature extraction gives out the image, It's size is further reduced by applying max-pooling technique. The image is now forwarded to fully integrated CNN layers. Character prediction based on 97.83% of the professional accuracy obtained by this model. The model is compiled by Adam's optimizer which is an extension of the stochastic gradient descent learning algorithm. This proposed model has found better accuracy and performance than older models.