arxivcs.CVcs.AI2026-07-15
Barnamala: Parameter-Efficient Handwritten Devanagari Recognition at Benchmark Saturation
We built a compact convolutional network (1.11 M parameters) for 46-class DHCD Devanagari recognition and reached 99.73%, the highest reported at 15.6x smaller than prior state-of-the-art. We have effectively reached the saturation point: every model tested, large teacher ensembl…