Stacking Diverse Deep Architectures for 62-Class Handwritten Character Recognition

This work presents an approach to improving the accuracy of handwritten character recognition (HCR) across 62 classes: the uppercase alphabet, the lowercase alphabet, and the Arabic numerals. A majority of the difficulties that arose during the process were a result of structurally similar characters such as 1/l/I, O/0, and upper/lower case similarities like W/w. We evaluated our approach on a Kaggle dataset where no prior model had exceeded 85% accuracy. We utilized a stacking ensemble of three diverse models: a ResNet-18, a CBAM CNN, and a hybrid CNN-ViT, combined with a logistic regression meta-learner trained on out-of-fold predictions with temperature-scaled probabilities. Training employed focal loss, online hard example mining (OHEM), and an augmentation pipeline including CutMix and elastic distortion. The stacked ensemble achieved an accuracy of 91.2%, outperforming the previous best submission which achieved 84.6%. Future work could incorporate real-world image sources and accurate character scaling to further improve accuracy on case-ambiguous pairs.