Improved Model Configuration Strategies for Kannada Handwritten Numeral Recognition

Handwritten numeral recognition has been an important area in the domain of pattern classification. The task becomes even more daunting when working with non-Roman numerals. While convolutional neural networks are the preferred choice for modeling the image data, the conception of techniques to obtain faster convergence and accurate results still poses an enigma to the researchers. In this paper, we present new methods for the initialization and the optimization of the traditional convolutional neural network architecture to obtain better results for Kannada numeral images. Specifically, we propose two different methods- an encoderdecoder setup for unsupervised training and weight initialization, and a particle swarm optimization strategy for choosing the ideal architecture configuration of the CNN. Unsupervised initial training of the architecture helps for a faster convergence owing to more task-suited weights as compared to random initialization while the optimization strategy is helpful to reduce the time required for the manual iterative approach of architecture selection. The proposed setup is trained on varying handwritten Kannada numerals. The proposed approaches are evaluated on two different datasets: a standard Dig-MNIST dataset and a custom-built dataset. Significant improvements across multiple performance metrics are observed in our proposed system over the traditional CNN training setup. The improvement in results makes a strong case for relying on such methods for faster and more accurate training and inference of digit classification, especially when working in the absence of transfer learning.