Test-Time Adaptation for Low-Resource Handwritten Character Recognition under Distribution Shift

This study presents the first systematic evaluation of test-time adaptation for low-resource handwritten character recognition, demonstrating that on-the-fly batch normalization (BN) recalibration recovers up to 3× the accuracy lost to noise and degradation without retraining, while state-of-the-art gradient-based adaptation methods catastrophically fail on a high-capacity pre-trained model. The method is evaluated on Yorùbá handwritten character recognition using two architectures, a custom convolutional neural network (YorùbáNet) and a fine-tuned ResNet-18, across synthetic corruptions including Gaussian blur, salt-and-pepper noise, and stroke thinning. Results show that BN-based TTA improves performance under intensity-based corruptions. Accuracy increases from 7.2% to 22.1% and from 19.9% to 44.9% for YorùbáNet and from 85.8% to 96.5% and from 92.4% to 96.9% for ResNet-18 under noise and thinning, with statistically significant gains (p < 0.001) and low computational overhead. Limited improvement is observed under Gaussian blur, indicating dependence on the type of distribution shift. Comparison with entropy minimization and self-supervised test-time training shows that these methods can improve weaker models but may reduce performance in stronger models, while BN-based adaptation remains stable. The results indicate that BN-based TTA improves robustness and calibration in low-resource Optical Character Recognition (OCR) without retraining.