StrikeNet — strike/clean word classifier

StrikeNet is the 79k-parameter CNN that pdf-strikethrough-detect uses to resolve pixel-ambiguous strikethrough cases on scanned document pages. A thin strike over an ascender-less word is pixel-identical to an ordinary glyph chain; geometry alone cannot separate them, so a learned model casts the deciding vote. On born-digital PDFs the package reads vector strokes and annotations directly and never invokes this model.

• Input: one standardized grey word crop, 32×160 (ink-positive, height-normalized). Produced by pdf_strikethrough.cnn.std_crop; the exact geometry is recorded in strike_verdict_cnn.meta.json and enforced at load time (cnn._check_geometry) so training and inference preprocessing cannot drift.
• Output: a single logit → sigmoid probability that the word is struck.
• Decision thresholds (in the meta): p_hi = 0.85, p_lo = 0.15. A word scoring ≥ p_hi is struck, ≤ p_lo is clean, and between them it is "unsure". Of the two thresholds, only p_hi decides what is reported: a candidate the CNN does not confirm at p_hi is not reported (a geometry-confident one whose crop is too small to score is kept), and p_lo only labels the rest clean or unsure. How the shipped values were chosen is not recorded (see Provenance); the training script sets a retrained model's thresholds from data: p_hi as a split-conformal threshold on held-out struck-word probabilities (a recall floor of 1 − alpha for crops like the validation set; see training/README.md), p_lo mirrored on the clean class. Override per call with ScanConfig.recall_first(cnn_p_hi=…) / precision_first(cnn_p_hi=…).

How it fits the pipeline

Scanned page → OCR words + morphological stroke geometry (lines.py) → StrikeNet adjudicates the auto/review tier crops → struck words, char-level partial resolution, struck-aware markdown. Full architecture: the project README.

Evaluation

Measured on the project's reproducible 10-document public regulatory-redline corpus (US Copyright Office, FDIC, CEQ ×2, EPA ×3, California CCPA/CPPA, Gretna LA development code; 55.2k struck words, each sha256-pinned). Reproduce with benchmarks/scanned_recovery.py.

Metric	Value	Source
Scanned-path strike recovery (RapidOCR)	96%	scanned_recovery.py, 3 docs / 24 pages / 2,170 known strikes
Scanned-path strike recovery (Azure DI)	96% with confidence_free(); 91% with the default DI calibration	same harness, Azure Document Intelligence words
Native vector detections independently confirmed by the flag signal	99.8%	confirmation_rate.py (context; native path, not this model)

The two scanned-path figures measure the whole pipeline (stroke geometry, OCR words and this model together), not the model on its own. Whether these documents overlap the model's training data is not recorded (see Provenance). No held-out precision/recall figure for the model itself exists yet: it needs labeled crops the model cannot have seen, such as from documents published after July 2026, or a retrain on a recorded labeled set.

Usage

The package runs on ONNX Runtime with no torch dependency. Download is digest-verified before the graph is ever loaded — a tampered host cannot swap the model.

import pdf_strikethrough as st

st.ensure_model(
    "https://huggingface.co/niles-liu/strikenet/resolve/main/strike_verdict_cnn.onnx",
    "fac2c51baaa75ee782196bdfe7452638cb48c7deddb21163b1ac6a0a72ae4457",
    meta_url="https://huggingface.co/niles-liu/strikenet/resolve/main/strike_verdict_cnn.meta.json",
    meta_sha256="4388b14715bfb1f51b56bb6c463d8f5c0847533316297890e70ac0b2234405d4",
)
assert "p_hi" in st.get_model_meta()          # thresholds + crop geometry now loaded
result = st.detect_pdf("scanned-redline.pdf", ocr=st.rapidocr_backend())

Provenance and training

The shipped weights cannot be reproduced from the package's repository. The ONNX the package ships (version: v3-corpus, also hosted at niles-liu/strikenet) is byte-identical in every release since the first (0.4.0, July 2026). It predates training/train_strikenet.py, which did not produce it. Its training set, train/validation split and training settings are not recorded in the package's repository or any release, and neither is how its thresholds (0.85 / 0.15) were chosen: the first release's code calls them guardrails, nothing more.

Training a replacement. training/train_strikenet.py trains StrikeNet on a labeled crop set the detector exports, sets p_hi / p_lo from held-out data, and records in the meta, under training, what the model was trained on: a digest of the labeled crops and their labels, the class counts, the validation size, the hyperparameters (the seed that draws the split among them) and the package and torch versions. The digest identifies the labeled set but cannot restore it, so keep the labeled directory. The failing-page → better-model loop is one command per step:

pdf-strikethrough detect scan.pdf --ocr rapidocr --dump-crops crops_out/   # export scored crops
# label crops_out/crops.jsonl: set each row's "label" to "struck" or "clean"
python training/train_strikenet.py crops_out/ -o model_out/ --epochs 40    # train + calibrate + ONNX

Full loop and label format: training/README.md.

Limitations

• Handwritten / freehand strikes are the main known gap: untested, and nothing records that the model saw any in training, so a wavy pen scribble or heavy cross-out may be missed.
• Horizontal, left-to-right text only. Vertical writing modes (CJK/Mongolian) and RTL strike axes are out of scope.
• The model only sees word crops the geometry stage flags as candidates; it does not itself find words on the page.

License

MIT, matching the parent package.