Legaldocuman Rfdetr

legaldocuman-rfdetr

RF-DETR-Base fine-tuned for handwritten signature detection in legal documents.
This model is the computer vision backbone of LegalDocuMan — a commercial document intelligence API for contract processing.

---

Model Details

Field	Value
Base architecture	RF-DETR-Base
Parameters	31.9M
Base checkpoint	Roboflow RF-DETR-Base (COCO pretrained)
Task	Single-class object detection
Class	signature
Input resolution	560px (square)
License	Apache 2.0

---

Training Data

Three open-source datasets were merged, deduplicated, and cleaned before training.
Cross-split contamination was verified via perceptual hashing — zero overlap between train, val, and test confirmed.

Dataset	Images	License
signatures-xc8up (Roboflow 100)	~2,800	CC BY 4.0
Signature Detector (TrueSign)	772	MIT
Signature Detection v3 (home)	2,145	CC BY 4.0

Final dataset size after cleaning: 3,996 images

---

Training Configuration

Parameter	Value
Augmentation preset	Document (9 transforms)
Augmentations	Perspective distortion, grid distortion, JPEG compression artifacts, Gaussian blur, Gaussian noise, random brightness/contrast, CLAHE, sharpening, horizontal flip
Input resolution	560px
Early stopping patience	20 epochs
Stopped at	Epoch 16
Best val mAP@50	0.9123 (epoch 11)
Framework	rfdetr · PyTorch · PyTorch Lightning
Hardware	NVIDIA RTX 4070 Ti Super 16GB

---

Evaluation

Evaluated on 168 held-out test images with zero contamination from training or validation sets.

Overall

Metric	Score
mAP@50	0.7924
mAP@50:95	0.820
Precision@0.50	0.9490
Recall@0.50	0.8020
F1@0.50	0.8696

By Signature Size

Size	AP	AR
Small	0.252	0.500
Medium	0.834	0.855
Large	0.828	0.845

Note on small signatures: Signatures smaller than ~32×32px after resizing have significantly lower recall.
For documents with initials or compact date-field signatures, human review is recommended.

---

Usage

from rfdetr import RFDETRBase
from PIL import Image

model = RFDETRBase(pretrain_weights="Mo-Awadalla/legaldocuman-rfdetr")

image = Image.open("contract_page.jpg")
detections = model.predict(image, threshold=0.45)
print(detections)

For PDF inputs, convert pages to images first using pdf2image:

from pdf2image import convert_from_path

pages = convert_from_path("contract.pdf", dpi=200)
for i, page in enumerate(pages):
    detections = model.predict(page, threshold=0.45)
    print(f"Page {i+1}: {detections}")

---

Intended Use

Designed for:

• Detecting the presence and location of handwritten ink signatures in legal contract pages
• Document intake pipelines processing PDF, DOCX, and scanned image inputs
• Execution status classification (executed vs. draft) as part of a broader document intelligence pipeline

Out of scope:

• Signature verification or authenticity determination
• Forgery detection
• Digital or electronic signature detection
• Handwriting recognition or transcription

---

Limitations

• Small signatures (initials, compact date-field signatures) have significantly lower recall (AP 0.252)
• Performance may degrade on scans below 150 DPI
• Not trained on non-Latin document layouts
• Should not be used as the sole decision-maker for high-stakes legal determinations without human review in the loop

---

Part of LegalDocuMan

This model is one component of the LegalDocuMan pipeline:

Upload PDF/DOCX/Image
        ↓
Text extraction (pdfplumber · python-docx · Tesseract OCR)
        ↓
Document type classification (MSA · SOW · NDA · PO · Amendment · License · Contract)
        ↓
Execution status (Regex NLP + RF-DETR visual signature detection)
        ↓
Vendor · Dates · Retention extraction
        ↓
Structured filename + PostgreSQL persistence

→ GitHub: Mo-Awadalla/LegalDocuMan

---

Attribution

Training data includes datasets licensed under CC BY 4.0.
Per license requirements:

• signatures-xc8up by Roboflow 100 — Roboflow Universe
• Signature Detection v3 by home — Roboflow Universe