Send your OCR only the regions that matter
Ultralytics YOLO finds fields, tables, signatures, and stamps and crops them for any OCR engine or VLM, trained on your documents in Ultralytics Platform and licensed for your product.
Trusted by the world's leading organizations
Built for teams that ship document AI
Ultralytics provides the models, tools, and license. You build and own the product, from OCR to review queues.
Document AI and IDP platforms
Add layout, form-field, and table detection to your extraction product, and fine-tune it on each customer's documents.
Construction and engineering software
Find and count symbols, fixtures, and rooms on blueprints, floor plans, and P&IDs for takeoff and estimating tools.
Identity, banking, and in-house teams
Check that signatures and stamps are present, locate ID cards, and mask sensitive regions on-premises before any OCR runs.
What published research shows
Peer-reviewed results from models built on Ultralytics YOLO.
How it works
From a scanned page to clean crops your OCR can read.
Bring in scans, PDFs, and photos
Render each PDF page or scan as an image, and tile large drawings so small symbols stay visible.
Classify and straighten each page
Sort pages by type, and find rotation and page corners so each one can be turned upright and flattened.
Detect and crop the regions
YOLO finds fields, tables, signatures, and stamps with exact coordinates, then crops each one for your OCR engine or VLM, or masks it.
Fine-tune on your documents
Label a few hundred representative pages in Ultralytics Platform, measure accuracy per class, and deploy to your own servers.
Everything YOLO can find on a page
Five YOLO tasks in one model family, trained on your pages and deployed the same way.
Try YOLO26 OCR
Drag and drop an image or PDF to read its text in real time
Why teams build on Ultralytics YOLO
Exact region detection on your own hardware at a fixed cost, with no per-page API fees.
Runs on CPU or GPU
Export to ONNX, OpenVINO, TensorRT, and CoreML, then run on your own servers, desktops, or capture devices.
Private by design
Run fully on-premises or offline, and mask card numbers, ID photos, and other sensitive regions before any page reaches a cloud OCR.
Boxes you can check
Outputs are coordinates with confidence scores, not generated text, so low-confidence fields go to review instead of being guessed.
Two ways to start
Train and deploy on Ultralytics Platform, or add YOLO to your own pipeline.
Ultralytics Platform
Upload sample pages, label fields and layout regions with AI-assisted annotation, train a custom YOLO model, and export it for your servers, all in one place.

Python
Fine-tune YOLO on the signature dataset or your own form fields, then get each field's class and location in a few lines. License YOLO to ship it in your own product.
from ultralytics import YOLO
model = YOLO("yolo26n.pt")
# Train on signatures, or your own form fields
model.train(data="signature.yaml", epochs=100)
result = model("contract.jpg")[0]
for box in result.boxes: # each field and its location
print(result.names[int(box.cls)], box.xyxy.tolist())Enterprise-grade security and compliance
Build and deploy with confidence. ISO 27001 certified, SOC 2 attested, with support for HIPAA compliance.
Frequently asked questions
No. YOLO finds and classifies regions such as fields, tables, signatures, stamps, and checkboxes, and crops them for any OCR engine or VLM to read.
You build and own the application. Ultralytics provides the YOLO models, Ultralytics Platform to train on your pages, and the commercial license. Your team adds OCR, validation rules, review queues, and integrations.
Yes, after fine-tuning on your pages. Render PDFs to images, tile large drawings so small symbols stay visible, and label a few hundred representative pages per new class. The Signature Detection dataset is a quick place to start.
It detects whether a signature, stamp, or ID field is present and where it is. Whether a signature is genuine or an ID is valid is decided in your verification layer, for example by e-signature validation or MRZ checks.
Yes. Export to ONNX or OpenVINO and run inference on your own servers, offline if needed. Train locally, or in Ultralytics Platform's US, EU, or AP data region.
Detect sensitive regions on your own hardware and mask them before any page goes to a cloud OCR. Object blurring in the Ultralytics package does this in a few lines.
Usually not. Recital 53 names classifying documents and turning unstructured data into structured data as narrow procedural tasks. Uses that decide credit, hiring, or life and health insurance pricing are high-risk under Annex III.
Ultralytics YOLO is open source under AGPL-3.0, so you can prototype freely. To ship it in a closed-source product or service, an Enterprise License removes the AGPL obligations, with no per-page fees. Enter your work email above to get a quote.
More use cases

Retail loss prevention

Visual inspection

Driver monitoring

Inventory management

Defect detection

Queue management
Your vision, realized
Join millions of developers building scalable solutions for smart use cases across industries. Annotate, train, and deploy — all from Ultralytics platform.



