archival-index-card-detector-v4

Archival Index Card Detector v4

A YOLO26n model trained to detect index cards in digitized archival document scans. This is a development version for validation before promoting to the NationalLibraryOfScotland/archival-index-card-detector repo.

Changes from v3

• Expanded training data: 1015 annotations (up from 905), including 110 human-reviewed images from the late-range Al-Allan scans (images 1127+)
• Better late-range recall: v3 had ~24% recall on images 1127+ due to a scan configuration shift (different background appearance). v4 specifically targets this gap.
• Background diversity: Added 20 new negative examples (card backs, blank pages) from the late range

Performance

Best epoch: 59/100 (early stopping at 79, patience=20).

Training Details

• Base model: YOLO26n (ultralytics)
• Training images: 1015 (811 train, 204 val)
• Positive examples: 542 (with bounding boxes)
• Negative examples: 473 (card backs, blank pages, etc.)
• Epochs: 79 (early stop), best at 59
• Hardware: Apple M1 Pro (MPS)
• Framework: ultralytics

Usage

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

model = YOLO(hf_hub_download(
    repo_id="davanstrien/archival-index-card-detector-v4",
    filename="model.pt"
))

results = model.predict("scan.jpg")

for result in results:
    for box in result.boxes:
        x1, y1, x2, y2 = box.xyxy[0].tolist()
        confidence = box.conf[0].item()
        print(f"Card at ({x1:.0f}, {y1:.0f}, {x2:.0f}, {y2:.0f}) conf={confidence:.2f}")

Status

Development/validation - pending comparison against v3 on late-range images before promotion to the NLS org repo.