Zhnivo/spread-classifier

Spread Classifier: two-page spread or single page

Decides whether a scan shows a two-page spread of a bound book or register, or a single page or sheet. Spreads can then be cut into two pages before layout analysis and text recognition.

Works in a pair with Zhnivo/spread-gutter-locator: this model finds the spreads, the gutter locator finds the line where each spread should be cut into two pages.

Released by Zhnivo, a non-profit project preserving the archival heritage of Belarus.

Model

Architecture	EfficientNet-B0 (torchvision) with a single-logit head
Input	RGB scan, longest side 512 px, centred on a gray (128) 512×512 square, ImageNet normalisation
Output	logit that the scan is a two-page spread
Recommended threshold	logit > 4.5 means spread (lower it to find more spreads at the cost of more false cuts)
Hardware and speed	runs on an ordinary CPU, no GPU needed: about 5 scans per second (around 17,000 an hour) on a laptop CPU (AMD Ryzen 7, 6 threads), including reading the scans
Files	model.safetensors, config.json, inference.py

Training data

Trained on several thousand scans of printed and handwritten books, registers and archival files, mostly from the 19th and 20th centuries: two-page spreads and single pages or sheets, including large tables and forms. Validation scans come from other books than the training scans.

Usage

pip install torch torchvision safetensors pillow numpy huggingface_hub

from huggingface_hub import snapshot_download
import sys

folder = snapshot_download("Zhnivo/spread-classifier")
sys.path.insert(0, folder)
from inference import load_model, predict

model, cfg = load_model(folder)
for z in predict(["scan.jpg"], model, cfg):
    print(z, "spread" if z > cfg["recommended_logit_threshold"] else "single")

Prepare the input exactly as inference.py does: no stretching, gray padding.

Limitations

• Built for scans and photos of archival books, registers and files. Other images are out of scope.
• Large single sheets folded or creased in the middle (wide tables) are the most common false spreads.
• The model only says whether to cut, not where.

License and attribution

MIT, free for any use. If you use the model, a link to the project zhnivo.com is appreciated.