Convolutional Neural Network Based Kannada-MNIST Classification

Before machine learning emerged, tasks that involve the recognition of handwritten characters, which include postcode recognition, document digitalization, and ancient character recognition, all require extensive human effort. Nowadays, various methods have been developed so that such tasks can be handled efficiently and accurately by computers. In 2018, the School of Computer and Information Science of Southwest University and the Research Institute of Yi Nationality at Guizhou University of Engineering Science jointly took lead in the research of ancient Yi language classification by using artificial intelligence technology. A model based on Convolutional Neural Networks (CNN) was developed and achieved a rather high accuracy. This triggered my great curiosity about handwritten character classification using machine learning methods.In 2019, the Kannada-MNIST (K-MNIST) dataset, a well-processed dataset of numerals in the Kannada language, was disseminated. We decided to make use of this dataset for our research in handwritten character classification. A CNN model was primarily developed due to its special architecture that makes it well suited for image classification tasks. This paper focuses on the establishment and experimentation of this CNN model and makes a detailed analysis. We also considered other methodologies such as Logistic Regression and Support Vector Machine for making comparisons, so that we can gain an insight into the performance of our model when compared to other machine learning methods. Through our experiment, the model achieved an accuracy of 98.77% over the testing set of K-MNIST, surpassing all the baselines, and it is effective for the classification of all categories of the dataset (0-9). Thus, we eventually concluded the strong capability of CNN models when performing classification tasks of handwritten characters.