HED-Tensor: Holistic Edge Detection
A TensorFlow/Keras implementation of Holistically-Nested Edge Detection (HED) with pretrained weights included. Use it as easily as VGG16 or ResNet!
🚀 Quick Start
Installation
pip install hed-tensor
Usage (3 lines!)
from hed_tensor import HED, detect_edges
model = HED() # Weights automatically load!
edges = detect_edges('your_image.jpg', model=model)
That's it! No need to download weights separately. 🎉
🆕 What's new in 2.0.0
- Retrained weights. The bundled
hed_bsds500.weights.h5was retrained from ImageNet VGG16 with the original authors' Caffe recipe: the full augmented HED-BSDS training set (28,800 images), 10 epochs (28,680 updates of 10 images), class-balanced loss summed over pixels, consensus labels, per-layer learning rates, SGD with momentum, learning rate 1e-6 divided by 10 every 10k updates. The 1.x weights came from a short run whose learning rate decayed to almost zero after 2 epochs, so their edge maps were mostly noise. - No unnecessary download.
HED()no longer downloads the ImageNet VGG16 weights (58 MB) when it loads HED weights; they are only used forHED(weights=None), i.e. to start training. - Errors are raised. A missing or incompatible weights file now raises an exception instead of printing a warning and silently returning an untrained model.
- Loss = original HED loss.
WeightedBCELoss/HEDLossnow match the Caffe implementation the weights were trained with (summed over pixels per image, averaged over images; labels 0/1). detect_edgesruns the model directly (faster, no re-tracing for each new image size);batch_detect_edgesandsave_edge_mapare now exported, as the examples below always assumed.- Requires TensorFlow ≥ 2.16 (Keras 3), which the weights file format needs.
📦 Features
✅ Pretrained weights included - No extra downloads needed
✅ Simple API - Just like using VGG16 or ResNet
✅ Multiple input formats - Works with file paths, PIL Images, or NumPy arrays
✅ Batch processing - Process multiple images efficiently
✅ TensorFlow 2.16+ / Keras 3 - Modern and fast
🎯 Examples
Basic Usage
from hed_tensor import HED, detect_edges
from PIL import Image
# Load model (weights auto-load)
model = HED()
# Detect edges
edges = detect_edges('image.jpg', model=model)
# Save result
Image.fromarray(edges).save('edges.png')
Using NumPy arrays
import numpy as np
from hed_tensor import HED, detect_edges
model = HED()
image_array = np.array(Image.open('image.jpg'))
edges = detect_edges(image_array, model=model)
Batch Processing
from hed_tensor import batch_detect_edges, HED
model = HED()
image_paths = ['img1.jpg', 'img2.jpg', 'img3.jpg']
edge_maps = batch_detect_edges(
image_paths,
model=model,
output_dir='output_edges'
)
Load Custom Weights
# Use your own trained weights
model = HED(weights='path/to/your/model.weights.h5')
# Or skip loading weights (ImageNet VGG16 backbone, for training)
model = HED(weights=None)
Side outputs
import tensorflow as tf
from hed_tensor import HED, preprocess_image
model = HED()
logits = model(preprocess_image('image.jpg'), training=True) # 5 side outputs + fused
side1, side2, side3, side4, side5, fused = [tf.sigmoid(l)[0, ..., 0].numpy() for l in logits]
📋 Requirements
- Python >= 3.9
- TensorFlow >= 2.16 (Keras 3)
- NumPy >= 1.19.0
- Pillow >= 8.0.0
🎓 Citation
If you use this package in your research, please cite the original HED paper:
@inproceedings{xie2015holistically,
title={Holistically-nested edge detection},
author={Xie, Saining and Tu, Zhuowen},
booktitle={Proceedings of the IEEE International Conference on Computer Vision},
pages={1395--1403},
year={2015}
}
📄 License
MIT License
👨💻 Author
Mohammad Saad Nathani
📧 saadnathani2005@gmail.com
🤝 Contributing
Contributions are welcome!
Made with ❤️ for the Computer Vision community
Metadata
Release files for hed-tensor 2.0.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| hed_tensor-2.0.0.tar.gz | 54.8 MB | Without provenance | Details for hed_tensor-2.0.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| hed_tensor-2.0.0-py3-none-any.whl | Python 3 | none | any | Details for hed_tensor-2.0.0-py3-none-any.whl |
Total release size: 109.5 MB
Release files / hed_tensor-2.0.0.tar.gz
| Download URL | hed_tensor-2.0.0.tar.gz |
|---|---|
| Size | 54.8 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
d71bf1e0806f9091eba0d4f194a9e81ba7168d19ed057e1a8718ca382ecf2e62
|
|
BLAKE2b-256 checksum How to use checksums |
d4a1e2fdf44d210fca3039b3405f9862bf9818d5e5826127b5d1cd0c9138ca41
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.11
|
Release files / hed_tensor-2.0.0-py3-none-any.whl
| Download URL | hed_tensor-2.0.0-py3-none-any.whl |
|---|---|
| Size | 54.8 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
8f22d716110ec4f440af6c1a99aa6706b49d30139a5fbb6dbafdec13a7f50c86
|
|
BLAKE2b-256 checksum How to use checksums |
8478e68e6bc020344b0db4d0d95f250871a32a75d8d7f3f8ec80029bd40877b5
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.12.11
|
Security for hed-tensor 2.0.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.