Multi-Character Recognition using EMNIST

In this paper, we present a model for handwritten multiple-character recognition. This model consists of a classifier which recognizes an individual character, and a series of image processing algorithms that extract individual character regions from a given image and feed it to the classifier. We make use of a Convolutional Neural Network (CNN)-a state-of-the-art solution to object recognition-for the task of classification, and train it on the recently published Extended MNIST (EMNIST) dataset, achieving an accuracy close to 90%. We then make use of Canny Edge Detection and dilation to segment the given image and feed it to the aforementioned classifier. The EMNIST dataset has the potential to become a standard benchmark in computer vision systems, however, limited literature on this dataset has been published as of now. Hence this paper seeks to further validate the dataset and give an indication of the potential performance achievable using the same.