davanstrien/britannica-illustrated-detector

Britannica illustrated-page detector (efficientvit_b1.r256_in1k)

Classifies scans of historical encyclopedia pages as illustrated (1) or not illustrated (0). Built as a fast, cheap replacement for a 2023 AutoTrain ViT-B/16@384 classifier, for pre-filtering and confirming illustration candidates across large scanned-book collections (e.g. the ~4 TB Internet Archive mirror of Encyclopaedia Britannica editions).

Usage

import timm, torch
from PIL import Image

model = timm.create_model(
    "hf-hub:davanstrien/britannica-illustrated-detector", pretrained=True
).eval()
data_cfg = timm.data.resolve_model_data_config(model)
tfm = timm.data.create_transform(**data_cfg, is_training=False)

probs = torch.softmax(model(tfm(Image.open("page.jpg").convert("RGB")).unsqueeze(0)), dim=-1)[0]
is_illustrated = probs[1].item() >= 0.982   # recommended threshold

Label mapping: index 0 = not illustrated, index 1 = illustrated (dataset ground-truth order).

• Recommended threshold: 0.982 on P(illustrated) — the best-F1 point on the held-out NLS split.
• Out-of-domain validation (2026-08-19, human-labelled): on 72 pages sampled from the Internet Archive Britannica mirror (different editions, scanners and OCR quality than the NLS training data), the model scored precision 1.000 / recall 0.872 at 0.982 — every miss was a near-threshold page at 0.91–0.98. For mirror-style harvesting, use threshold 0.90 (F1 1.0 on the labelled sample; negatives max 0.878, positives min 0.913).
• Input: 256×256, RGB (grayscale scans are replicated to 3 channels by the transform).

Results (same held-out split: 515 pages, 24 illustrated, seed 42)

model	params	input	PR-AUC	best-F1 (swept thr)	acc @ thr	confusion (TN/FP/FN/TP)
AutoTrain ViT-B/16 (baseline)	86 M	384²	1.000	1.000 @ 0.941	1.000	491 / 0 / 0 / 24
timm mobilenetv4_conv_medium (e500_r256_in1k)	8.4 M	256²	0.961	0.933 @ 1.000	0.994	491 / 0 / 3 / 21
timm efficientvit_b1.r256_in1k (this model)	7.5 M	256²	1.000	1.000 @ 0.982	1.000	491 / 0 / 0 / 24

Honest caveats:

• Both the baseline and this model saturate the split (perfect ranking of all 24 positives). At this validation scale the two models cannot be distinguished on quality — the case for this model is equal measured quality at ~11× fewer parameters and 9–37× cheaper inference (384²→256² input, 86M→7.5M params), not superiority.
• The baseline was trained in 2023 on the same NLS dataset via AutoTrain with an unknown split, so its rows above may be mildly optimistic (possible train/eval overlap from its original split); this model's split (seed 42, stratified 80/20) is disjoint from its own training data by construction.

Training

• Data: NationalLibraryOfScotland/encyclopaedia_britannica_illustrated (2,573 pages; 2,451 not-illustrated / 122 illustrated). CC0. Read directly from parquet (pyarrow).
• Recipe: timm.create_model(..., pretrained=True, num_classes=2), 12 epochs, batch 64, AdamW 3e-4 (cosine, ~3% warmup, wd 0.05), class-weighted cross-entropy with sqrt inverse-frequency weights (≈[0.22, 4.47]) for the 20:1 class imbalance.
• Augmentation (light — book pages have a canonical orientation): RandomResizedCrop 256 (scale 0.65–1.0), rotation ≤5°, brightness/contrast jitter 0.2; no flips.
• Split: stratified 80/20, seed 42. Epoch-12 checkpoint selected on val PR-AUC.
• Trained on HF Jobs (a10g-small, ~5 min wall-clock).

Deployed at scale

This model was run over every page of all ~1,219 scanned volumes of the Britannica mirror (~1,007,000 pages): 117,390 illustrated pages were extracted and published (with full page-level scores) in davanstrien/britannica-illustrated-pages.

Production thresholds (validated on 347 human-labelled pages across two independent samples): 0.90 for text-dense pages, 0.982 for low-text pages.

Intended use

Pre-filter / confirmation step for finding illustrated pages in historical encyclopedia and similar scanned-book collections (where "few words in djvu.xml" is a cheap first-stage signal and this model is the second stage). Not intended as a general illustration detector for non-page documents, modern books, or colour photography; validate the threshold on a sample of your own collection before bulk use.

Replaces

This model replaces davanstrien/autotrain-encyclopaedia-illustrations-blog-post-3327992158 — same task, 9–37× cheaper inference (metrics in the table above; lineage auditable via the matching metrics.json in the training repo).

⚠️ Migration note: that baseline's config.json id2label ("0": "illustrated") is inverted relative to the dataset convention used here. Statistical orientation checking on 515 held-out pages shows its logits actually follow the dataset convention (PR-AUC 1.000 as-read vs 0.025 if you trust its id2label). If you are migrating, keep the raw class indices and re-map explicitly — do not consume the old config's label names.