BanglaCLR: Structure-Aware Contrastive Representation Learning for Bengali Character Recognition

Bengali characters often differ by a short stroke, dot, or loop, making recognition sensitive to both shape and stroke layout. BanglaSAN encodes appearance and structure in separate branches, exchanges information between them through bidirectional cross-attention, and classifies the fused representation against graph-refined prototypes. Its contrastive loss assigns larger penalties to confusable negative classes. Because the 49 classes in BCRD each contain only one distinct glyph image, we generate non-overlapping degraded samples for training, validation, and testing. We compare seven model configurations and assess significance, calibration, robustness, and explanation faithfulness. BanglaSAN obtains 95.22% test accuracy and 2.9% expected calibration error. It outperforms the single-branch CNN and structure-only baselines, although several appearance-based ablations have statistically comparable accuracy. Rotation, blur, and occlusion tests favor the proposed model at high severity. Appearance-deletion experiments also show that the structure branch carries class information on its own. Gaussian noise is the main exception: corrupt edges sharply reduce the benefit of the structure pathway.