Feature-Level Class-Adaptive Zoning Framework for Recognition of Kannada Stone Inscription Character Images

Recognition of characters from historical stone inscriptions is a fundamental problem in digital epigraphy, complicated by severe surface degradation, limited annotated data, and intrinsically imbalanced class distributions. In Kannada stone inscriptions, character frequencies follow a long-tailed distribution, where rare yet epigraphically significant characters appear very infrequently. While existing imbalance-handling strategies primarily operate at the data or classifier level, imbalance-induced bias in feature construction has received limited attention. This paper proposes β _c -based Density-Modulated Zoning (β _c -DMZ), a featurelevel, class-adaptive zoning framework for imbalanced Kannada stone inscription character recognition. The proposed approach embeds class distribution information directly into handcrafted zoning features through a class-dependent scaling mechanism derived from inverse class frequency and modulated by local foreground density. This formulation selectively enhances discriminative stroke regions for minority classes while preserving feature dimensionality, classifier architecture, and interpretability. Importantly, the method does not modify the learning objective, loss function, or training data distribution, as class-adaptive scaling is applied solely during feature construction. The proposed framework is evaluated on a severely imbalanced Kannada stone inscription dataset using a strictly nested cross-validation protocol to ensure unbiased performance estimation. Experimental results demonstrate that β _c -DMZ consistently outperforms baseline zoning features and classifier-level cost-sensitive learning, yielding average of 90.20% in macro-averaged recall and 90.44% in macro-averaged F1-score, with statistically significant gains (p < 0.05) and reduced inter-fold performance variability. Performance improvements are particularly pronounced for rare and epigraphically important character classes. Analysis based on confusion matrices and zone-level visualizations indicates that the proposed framework strengthens zones corresponding to meaningful character strokes. These findings show that introducing imbalance awareness straight into feature creation can reliably and interpretably identify imbalanced characters in historical Kannada stone inscriptions.