T3.7 Multi-Script Indic Handwriting Recognition
Two-stage CNN pipeline: ScriptRouter (4-class) → ScriptCNN (per-script).
| Script | Classes | Top-1 | Top-5 | Macro F1 |
|---|---|---|---|---|
| Router | 4 | 99.92% | — | — |
| Devanagari | 46 | 99.48% | 99.99% | 99.41 |
| Tamil | 156 | 97.30% | 99.77% | 95.96 |
| Bengali | 84 | 93.02% | 98.79% | 93.12 |
| Telugu | 6 | 98.89% | 100.0% | 98.88 |
E2E CPU latency: 10.21 ms. Trained from scratch, no transfer learning.
Usage
from inference import load_pipeline, preprocess, predict
import os
os.environ["HF_MODEL_REPO"] = "dhruv10050/t3-7-indic-recognition"
pipeline, scripts = load_pipeline("./checkpoints")
Files
router_best.pth— ScriptRouter (4-class CNN)devanagari_best.pth— Devanagari classifier (46 classes)tamil_best.pth— Tamil classifier (156 classes)bengali_best.pth— Bengali classifier (84 classes)telugu_best.pth— Telugu classifier (6 vowel classes)