International Journal of Scientific Research and Engineering Development

Automatic recognition of handwritten characters remains a fundamental challenge in computer vision, with practical relevance across document digitization, postal automation, banking operations, and educational assessment. This paper presents a comparative study of two deep learning models—Multi-Layer Perceptron (MLP) and Convolutional Neural Network (CNN)—evaluated on the Extended Modified National Institute of Standards and Technology (EMNIST) dataset. Input images of 28x28 grayscale pixels undergo per-pixel normalization and channel reshaping prior to training. The CNN employs stacked convolution and max-pooling layers to capture spatial hierarchies inherent in handwritten glyphs, augmented by Dropout regularization to improve generalization. Experimental results demonstrate that the CNN achieves approximately 86-88% validation accuracy within 20 training epochs with stable convergence and minimal overfitting, outperforming the MLP baseline by nearly 7-8 percentage points. These findings confirm the superiority of spatially-aware architectures for the diverse writing styles present in the EMNIST benchmark.

Atharv Kadam,Om Bhosale, Vinayak Dhulubulu, Kedar Teke, Apurva Mohite,"Handwritten Character Recognition Using Deep Learning on EMNIST Dataset" International Journal of Scientific Research and Engineering Development, V9(2): Page(2120-2127) Mar-Apr 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.