LDED

Lightweight Document Edge Detection — real-time corner detection in the browser.

---

Why LDED

Mobile document scanning requires fast and accurate detection of document boundaries. Classical computer vision approaches are fragile under varying lighting conditions, cluttered backgrounds, and perspective distortion. At the same time, while both (iOS) and (Android) already ship with operating-system-level computer vision and machine learning capabilities for tasks such as edge or document detection, these native capabilities are generally not accessible from within the browser. Web applications therefore cannot rely on the optimized on-device models that native apps can use.

This creates a clear gap: existing deep learning models for document boundary detection are typically too large or too computationally expensive to run in real time in a browser environment, especially on end-user mobile devices. In practice, there are currently no broadly available browser-first models that combine the necessary accuracy, speed, and device-level efficiency for this task.

LDED addresses this gap by providing a neural network compact enough to run entirely client-side via WebAssembly. This enables real-time document boundary detection directly in the browser, without server roundtrips, without exposing sensitive images to external infrastructure, and without the operational cost of backend processing.

Highlights

• Real-time document corner detection (< 100 ms) directly in the browser
• Privacy-preserving — all inference happens on-device, no data leaves the client
• Trained on diverse datasets (DocCorner, MIDV-500/2020, SmartDoc-2015) with heavy augmentation
• 3-phase training curriculum: Backbone Freeze → Fine-Tuning → Quantization-Aware Training
• Targets > 92 % corner precision (PCK@5) with < 5 MB model size

Use Cases

• Browser-based document scanners
• Mobile receipt and invoice capture
• ID card and passport boundary detection
• Drop-in integration into existing web applications

Results

Training Curves (TensorBoard)	Evaluation Metrics

Benchmark

Metric	LDED (ours)	Target	SmartDoc Baseline
PCK@5 ↑	93.0 %	> 92 %	—
PCK@10 ↑	98.2 %	> 97 %	—
IoU ↑	97.4 %	> 90 %	—
NME ↓	0.39 %	< 3.0 %	—
Val Loss ↓	0.0026	—	—

↑ = higher is better, ↓ = lower is better. Results on validation split. Baseline columns to be filled after final evaluation.

Model Card

Property	Value
Architecture	MobileNetV3-Small + FPN-Lite + Heatmap Head
Input size	512 × 512 px
Parameters	~2.5 M
Model size (FP32)	~10 MB
Model size (INT8)	< 5 MB
Output	4 corner points (x, y) + confidence
Framework	PyTorch 2.x → ONNX

Training Configuration

Setting	Value
Optimizer	AdamW
Learning rate	1e-3 (phase 1), 1e-4 (phase 2)
Scheduler	CosineAnnealingLR
Batch size	16
Epochs	50 (phase 1) + 30 (phase 2)
Loss	Adaptive Wing + BCE + Coordinate
AMP	Enabled (CUDA)
Augmentation	Perspective, Brightness/Contrast, MotionBlur, JPEG, GaussNoise

Datasets

Dataset	Raw Samples	× Augmentation	Effective Samples	Split
DocCorner (HuggingFace)	~4,000	×3	~12,000	train / val / test
MIDV-500/2020 (synthesized)	~3,000	×3	~9,000	train / val
Mendeley Corner	~1,100	×3	~3,300	train / val
Total	~8,100		~24,300	—

Inference

Runtime	Device	Latency
ONNX Runtime Web (WebGPU)	Desktop browser	~60 ms
ONNX Runtime Web (WASM)	Desktop browser	~500 ms
ONNX Runtime Web (WebGPU)	iPhone (mobile)	~237 ms
ONNX Runtime Web (WASM)	iPhone (mobile)	~500 ms

Contents

• Architecture
• Training Guide
• Technical README

---

CAS project work by Simon Fuchs & Raphael Fuchs, 2026