Img2Line AI Service

This project is trained with customed dataset: flatart borderless

TEED: Tiny and Efficient Edge Detection

TEED is a research project for fast, lightweight, and accurate edge detection in images, based on a custom neural network architecture. It includes tools for data preprocessing, model training, evaluation, and post-processing, supporting both standard and custom datasets. The project also provides utilities for converting between raster (PNG) and vector (SVG) formats, leveraging the vtracer library and rsvg-convert.

Features

• High-Performance Edge Detection: Fast inference with precision, recall, and F1 metrics logged to TensorBoard
• Flexible Training: Train from scratch or fine-tune from pretrained weights; configurable loss functions and learning schedules
• Data Preprocessing: Convert input PNGs → SVGs → resize to 1024×1024 → convert back to RGB PNGs
• Post-Processing: Vectorize PNG results to SVGs; rasterize SVGs to PNGs for downstream applications
• Evaluation Suite: Compute precision, recall, and F1 scores with visual outputs for validation

Installation

Python Dependencies (Python ≥ 3.7)

pip install -r requirements.txt
pip install vtracer

System Packages

macOS:

brew install librsvg

Ubuntu/Debian:

sudo apt-get install librsvg2-bin

Windows: See librsvg Windows builds

Project Structure

├── checkpoints/         # Pretrained & custom model weights
├── data/                # Preprocessed images for training & testing
├── input_data/          # Raw PNGs for processing
├── output_data_png/     # Post-processed PNG outputs
├── output_data_svg/     # Post-processed SVG outputs
├── result/              # Inference results
├── ted.py               # TEED model definition
├── loss2.py             # Training loss functions
├── dataset.py           # Dataset loader and transforms
├── preprocessor.py      # PNG↔SVG conversion and resizing pipeline
├── main.py              # CLI for preprocessing, training, testing, and post-processing
└── utils/               # Helper modules (image I/O, activation functions, metrics)

Quick Start

1. Add your raw PNGs to input_data/

2. Run the end-to-end pipeline:

   python main.py --choose_test_data -1

3. Retrieve outputs:

   • Edge maps: result/
   • SVG vectors: output_data_svg/
   • Rasterized PNGs: output_data_png/

Evaluation

Launch TensorBoard to view training and validation metrics:

tensorboard --logdir runs/

Metrics logged: precision, recall, and F1 score per epoch

License

This project is released under the MIT License. See LICENSE for details.

Acknowledgements

• vtracer for raster-to-vector conversion
• librsvg for SVG rasterization
• Research by Xavier Soria Poma and contributors

For questions or contributions, please open an issue or submit a pull request.