LDED: Detecting Document Corners in Real Time — Directly in the Browser

When capturing a document with a smartphone, users expect the frame to fit immediately and cleanly. This edge detection has been a gap in the browser until now — and LDED (Lightweight Document Edge Detection) closes it.

Background

Classical image processing detects document edges through contrasts and lines. This works in the lab, but not in reality: changing light, cluttered backgrounds, and perspective distortion quickly push such methods to their limits.

iOS and Android have long since brought optimized models at the operating system level — but these are generally not accessible from a web application. And the available deep learning models are too large or too computationally intensive to run in real time on a mobile browser. The usual workaround: the image is sent to a server. This means the very document that one wants to protect leaves the device — ID card, invoice, contract.

Solution

LDED is a neural network small enough to run entirely client-side. It provides not just a mask, but directly the four corner points plus a confidence value:

Training was conducted in a three-stage curriculum — frozen backbone, fine-tuning, then Quantization-Aware Training — on around 8,100 raw images from DocCorner, MIDV-500/2020, and the Mendeley-Corner dataset, expanded to approximately 24,300 examples via augmentation.

Result

On the validation split, LDED consistently exceeds the set target values:

The picture is more nuanced regarding speed:

The project is explicitly marked as “in development,” and the WASM path remains the next optimization challenge.

Typical use cases:

• browser-based document scanners,
• receipt and invoice capture on mobile devices,
• detection of ID cards and passports,
• integration into existing web applications without an additional backend.

Why We Share This

LDED was developed as a CAS project within our team. We decided to make the code, architecture description, training guide, and a working browser demo publicly available — as a contribution to research and education, not as a finished product.

Both can be tried directly: The Browser Demo detects document corners live via camera or uploaded image — the processing remains on your device. The model, training code, and documentation are available in the Repository on GitHub.

Content-wise, this project aligns with a stance that runs through our work: Sensitive data should be processed where it is generated. With ExeoAI, we pursue the same principle at the enterprise level — sovereign, local analysis of sensitive documents instead of processing in foreign infrastructure. LDED shows that the approach also works at the other end of the chain: already at capture, in the browser, on the user’s device.