doclayout-yolo-indic

DocLayout-YOLO-Indic

Real-time document layout detection for 12 Indic scripts, built on DocLayout-YOLO (YOLOv10-m + GL-CRM).

This repository accompanies an M.Tech dissertation (Vignesh P, BITS Pilani WILP). It contains four model checkpoints that together form a controlled ablation study — and its central result is an honest negative finding: the two proposed adaptation techniques did not improve over a simple baseline. That finding, proven and explained, is the main scientific contribution.

---

1. What problem does this solve?

Document Layout Analysis (DLA) is the first step in digitising any document: before OCR can read the text, a model must find the structure — which regions are paragraphs, headings, tables, figures, lists, and so on. Every downstream stage inherits the errors of this step.

Fast, modern DLA detectors are trained almost entirely on English and Chinese documents and generalise poorly to Indic scripts, because those scripts are visually different:

• Shirorekha — the continuous horizontal headstroke of Devanagari, Bengali and Gurmukhi — looks like a table border to a Latin-trained model.
• Conjunct consonants and stacked vowel marks (matras) break line-height assumptions.
• Urdu is written right-to-left.
• Real pages mix scripts and contain rare region types (sidebars, pull-quotes).

Over 1.4 billion people use Indic languages, yet no public real-time detector had been demonstrably adapted and evaluated across the major scripts. This project builds one — and rigorously tests whether its own adaptation ideas actually help.

---

2. The model

• Architecture: DocLayout-YOLO = YOLOv10-m backbone + GL-CRM (Global-to-Local Controllable Receptive Module) + PAN-FPN neck + dual (one-to-many / one-to-one) detection heads.
• Parameters: ~19.97 M (20.01 M at 42 classes).
• Input: 1024×1024 (also evaluated at 640 and 1280).
• Vision-only: it does not read text or depend on OCR — it responds purely to visual structure.
• Speed: >100 FPS on an NVIDIA A100 — real-time.

---

3. The checkpoints (a controlled ablation)

The four files differ only in what training happened before the final fine-tuning on the IndicDLP benchmark. Because everything else is held identical, comparing them isolates the effect of each proposed contribution — this is what licenses causal, not correlational, conclusions.

Which to use: for inference, use config_A_reported_42cls.pt — it is the best model and the one the dissertation reports. (It is a stripped, deployment-ready checkpoint, ~41 MB; the others retain optimizer state and are ~242 MB.)

The two class ontologies

• 9 classes (intermediate stages): text_body, headline, table, figure, caption, advertisement, sidebar, pull-quote, decorative-frame — a coarse, script-agnostic set used for synthetic pretraining and self-training.
• 42 classes (final): the full IndicDLP ontology, grouped into seven families — body text, heading hierarchy, list hierarchy, figures/tables/formulae, Q&A structure, navigation/metadata, and publishing furniture.

The detection head is re-shaped between stages (27 → 9 → 42 classes); the backbone and neck weights transfer across all stages, and only the final classifier changes width.

---

4. Headline results (reported model, config_A)

• 0.525 [email protected] and 0.364 mAP@[0.5:0.95] on the 11,633-image IndicDLP test set (180,039 instances).
• Consistent across all 12 scripts — per-script mAP@[.5:.95] spans only 0.304–0.426, with no script failing catastrophically. (Best: Assamese 0.426; lowest: English 0.299.)
• >100 FPS on A100 — real-time preserved.
• Precision 0.590 / Recall 0.506 overall.

Scalability (accuracy vs. speed)

---

5. The key finding (a rigorous negative result)

Neither proposed contribution improved over plainly fine-tuning the public checkpoint. Config A (no synthetic, no self-training) is the best model — by +11.2 mAP on the test set over Config B.

This is statistically significant across all 12 scripts:

• Wilcoxon signed-rank: p < 0.001 (p = 0.00049)
• Paired t-test: p < 10⁻¹⁰
• Cohen's d = 7.37 (a very large effect)
• All 12/12 scripts improved under Config A; 95% CI on the gain [0.096, 0.114].

Mechanism — catastrophic forgetting. The validation→test gap widens from 2.6 to 11.2 mAP under synthetic pretraining — the signature of damaged generalisation, not mere no-gain. The synthetic corpus (23 templates, 9 classes) is narrower than the base model's original DocSynth-300K pretraining; 15 epochs to 0.985 synthetic-validation over-specialised the backbone, and the coarse 9-class intermediate ontology erased fine distinctions the 42-class task then had to relearn from only 12,082 fine-labelled images.

In short: the techniques failed, but the science succeeded — a falsified hypothesis, proven under controlled conditions and mechanistically explained, that saves others from the same mistake.

Note on config_B and config_C

These are provided for transparency and reproducibility of the ablation — they are not recommended for deployment, as they under-perform Config A. They let others verify the negative result independently.

---

6. Known limitations

• Hierarchical-depth confusion is the dominant error. The model localises headings/lists but cannot infer nesting depth — recall collapses with depth (section-title 0.66 → sub 0.23 → subsub 0.01). This is architectural (a single-shot detector classifies from local appearance; depth is relational) and not Indic-specific — it affects English identically.
• Urdu is untested. Urdu was synthesised and trained on, but the IndicDLP benchmark contains no Urdu, so right-to-left performance is unevaluated.
• An auxiliary script-classification head was designed but not implemented (future work).
• Results use a 12,082-image training subset (compute-bounded); the reported mAP is therefore a lower bound, not a ceiling.

---

7. Usage

from huggingface_hub import hf_hub_download
from doclayout_yolo import YOLOv10

# download the reported model from this repo
ckpt = hf_hub_download("VigneshPR/doclayout-yolo-indic", "config_A_reported_42cls.pt")
model = YOLOv10(ckpt)

# run on a document page
result = model.predict("page.jpg", imgsz=1024, conf=0.25)[0]
result.plot()                      # annotated image (numpy array, BGR)
print(len(result.boxes), "regions")
for c in result.boxes.cls.tolist():
    print(result.names[int(c)])

Install the runtime:

pip install git+https://github.com/opendatalab/DocLayout-YOLO.git

---

8. Training setup

• Platform: Google Colab Pro+ (no institutional cluster). Training on A100 40 GB; evaluation on Tesla T4 to conserve compute.
• Data engineering under a 15 GB Drive limit: partial-tar extraction reduced the 42 GB IndicDLP download to ~5 GB; bulky image sets kept on ephemeral local disk, only labels/checkpoints persisted.
• Fine-tuning: ~30 epochs per config, batch 16 @ 1024, AMP, identical settings across all three ablation configs.
• Self-training (CBST): class-balanced pseudo-labelling on the unlabelled BaDLAD Bengali pool, with provably disjoint splits (leakage-audited: 20,365 − 10,000 = 10,365).

---

9. Related resources

• Synthetic corpus: VigneshPR/indicsynth-150k — the 150,000-page SA-MCB dataset used for pretraining.
• Benchmark (working mirror): VigneshPR/IndicDLP — mirror of the IndicDLP benchmark (original authors: IndicDLP, ICDAR 2025); credit and licence belong to them.

---

10. Citation

If you use these models or the finding, please cite the dissertation:

@mastersthesis{vignesh2026doclayoutindic,
  title  = {DocLayout-YOLO-Indic: Cross-Script Document Layout Analysis for Indic Scripts
            using Synthetic Pretraining, Self-Training and Controlled Ablation},
  author = {Vignesh P},
  school = {BITS Pilani (WILP)},
  year   = {2026}
}

Built on DocLayout-YOLO (YOLOv10-m + GL-CRM).