license: other license_name: docforensics-dual license_link: LICENSE.md tags:

• document-layout-analysis
• object-detection
• onnx
• document-ai
• ocr
• forms library_name: onnx pipeline_tag: object-detection

docforensics-layout-s

A 8.8 MB document layout detector for damaged, scanned and photographed pages. Runs in a browser through onnxruntime-web, on a server through onnxruntime-node, or anywhere ONNX runs. No GPU.

It detects fifteen region types including form fields and checkboxes, and it reports eight page-condition signals (blur, skew, tearing, bleed-through, handwriting, and whether the page is a scan carrying an invisible OCR layer) that no comparable detector produces.

What it is good at

On held-out degraded pages it reaches 0.449 mAP@50-95 where YOLOv8n-DocLayNet reaches 0.194, at 71% of that model's size. Across 120 held-out degraded pages the baseline returned no regions at all on 14 of them; this model returned more regions on 104.

| | this model | YOLOv8n-DocLayNet | |---|---:|---:| | size | 8.77 MB | 12.27 MB | | degraded mAP@50-95 | 0.449 | 0.194 | | clean mAP@50-95 | 0.477 | 0.660 | | form fields | AP 0.670 | no such class | | page-condition signals | 8 | none |

What it is not good at

Clean, born-digital pages rendered at high quality. YOLOv8n-DocLayNet is better there, 0.660 against 0.477, and if that is your input you should use it. The deficit is uniform across all ten shared classes, which reflects parameter count rather than a class-specific weakness.

Picture-heavy magazine layout. On a 1959 illustrated magazine page the baseline found 17 regions to this model's 10, separating advertisement panels and captions more finely.

Checkboxes at 640 px. A checkbox is roughly 6.5 px at this input size, below the finest anchor stride, so AP is 0.153. Read them from the PDF's AcroForm widgets instead when the page is born-digital.

Use it

import cv2, numpy as np, onnxruntime as ort

CLASSES = ["text","heading","list","table","picture","caption","formula",
           "header","footer","footnote","seal","form","field","checkbox","signature"]

sess = ort.InferenceSession("docforensics-layout-s.onnx",
                            providers=["CPUExecutionProvider"])

def letterbox(img, side=640):
    h, w = img.shape[:2]
    s = side / max(h, w)
    small = cv2.resize(img, (int(w * s), int(h * s)))
    canvas = np.full((side, side, 3), 114, np.uint8)
    canvas[:small.shape[0], :small.shape[1]] = small
    return canvas

img = letterbox(cv2.imread("page.png"))
x = cv2.cvtColor(img, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
x = ((x - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225]).transpose(2, 0, 1)[None]

det, signals, maps = sess.run(None, {"images": x.astype(np.float32),
                                     "evidence": np.zeros((1, 4), np.float32)})

# det is [1, 19, 8500]: rows 0-3 are cx,cy,w,h in 640px space, rows 4-18 are
# per-class scores. Apply your own confidence threshold and NMS.
boxes, scores = det[0, :4].T, det[0, 4:].T
keep = scores.max(1) > 0.18
print(f"{keep.sum()} candidate regions before NMS")

# signals is [1, 8]: skew, blur, noise, bleed, warp, tears, handwriting, native
print(dict(zip(["skew","blur","noise","bleed","warp","tears","hand","native"],
               signals[0].round(3))))

Outputs are YOLOv8-shaped on purpose, so any existing YOLOv8 decode and NMS works unchanged.

Node

detect() takes a canvas or raw pixels, so a server needs no DOM shim:

import { load, detect } from '@canwork/docforensics';
import ort from 'onnxruntime-node';
import sharp from 'sharp';

const model = await load('docforensics-layout-s', ort);
const { data, info } = await sharp('page.png')
  .removeAlpha().raw().toBuffer({ resolveWithObject: true });
const { regions, signals } = await detect(model, {
  data, width: info.width, height: info.height,
});

Node uses the CPU execution provider and is about three times faster than the browser's WebAssembly build. The output is identical either way.

Inputs

images is [1, 3, 640, 640], RGB, ImageNet-normalised, letterboxed with fill value 114. evidence is [1, 4] and may be all zeros; it carries optional deterministic page flags and is reserved for future use.

How it was trained

68k DocLayNet pages under a synthetic degradation pipeline covering skew, blur, sensor noise, bleed-through, perspective warp, physical tearing and aged-paper fading, plus 2,184 form pages labelled directly from AcroForm widget geometry and 6,821 pages labelled by a deterministic vector extractor. Training data is CDLA-Permissive-1.0 (DocLayNet) and US federal public domain (IRS forms, NASA technical reports). No proprietary or unlicensed material was used.

Licence

Dual-licensed, AGPL-3.0 or commercial. See LICENSE.md.

Published by Ginexys, a d/b/a of Canworks, LLC. Commercial licensing: [email protected]. Licence terms and compliance: [email protected].

The weights are separable from the surrounding pipeline and were trained entirely on permissively licensed and public-domain data.

Citation

See CITATION.cff.

A technical report documents the evaluation methodology, including a measured decorrelation between detection mAP and downstream extraction quality. It is not published yet; write to [email protected] if you need it before then.