Image Deduplicator (imagededup)

imagededup is a python package that simplifies the task of finding exact and near duplicates in an image collection.

This package provides functionality to make use of hashing algorithms that are particularly good at finding exact duplicates as well as convolutional neural networks which are also adept at finding near duplicates. An evaluation framework is also provided to judge the quality of deduplication for a given dataset.

Following details the functionality provided by the package:

• Finding duplicates in a directory using one of the following algorithms:
  • Convolutional Neural Network (CNN) - Select from several prepackaged models or provide your own custom model.
  • Perceptual hashing (PHash)
  • Difference hashing (DHash)
  • Wavelet hashing (WHash)
  • Average hashing (AHash)
• Generation of encodings for images using one of the above stated algorithms.
• Framework to evaluate effectiveness of deduplication given a ground truth mapping.
• Plotting duplicates found for a given image file.

Detailed documentation for the package can be found at: https://idealo.github.io/imagededup/

imagededup is compatible with Python 3.8+ and runs on Linux, MacOS X and Windows. It is distributed under the Apache 2.0 license.

📖 Contents

• Installation
• Quick Start
• Benchmarks
• Contribute
• Citation
• Maintainers
• License

⚙️ Installation

There are two ways to install imagededup:

• Install imagededup from PyPI (recommended):

pip install imagededup

• Install imagededup from the GitHub source:

git clone https://github.com/idealo/imagededup.git
cd imagededup
pip install "cython>=0.29"
python setup.py install

🚀 Quick Start

In order to find duplicates in an image directory using perceptual hashing, following workflow can be used:

• Import perceptual hashing method

from imagededup.methods import PHash
phasher = PHash()

• Generate encodings for all images in an image directory

encodings = phasher.encode_images(image_dir='path/to/image/directory')

• Find duplicates using the generated encodings

duplicates = phasher.find_duplicates(encoding_map=encodings)

• Plot duplicates obtained for a given file (eg: 'ukbench00120.jpg') using the duplicates dictionary

from imagededup.utils import plot_duplicates
plot_duplicates(image_dir='path/to/image/directory',
                duplicate_map=duplicates,
                filename='ukbench00120.jpg')

The output looks as below:

The complete code for the workflow is:

from imagededup.methods import PHash
phasher = PHash()

# Generate encodings for all images in an image directory
encodings = phasher.encode_images(image_dir='path/to/image/directory')

# Find duplicates using the generated encodings
duplicates = phasher.find_duplicates(encoding_map=encodings)

# plot duplicates obtained for a given file using the duplicates dictionary
from imagededup.utils import plot_duplicates
plot_duplicates(image_dir='path/to/image/directory',
                duplicate_map=duplicates,
                filename='ukbench00120.jpg')

It is also possible to use your own custom models for finding duplicates using the CNN method.

For examples, refer this part of the repository.

For more detailed usage of the package functionality, refer: https://idealo.github.io/imagededup/

⏳ Benchmarks

Update: Provided benchmarks are only valid upto imagededup v0.2.2. The next releases have significant changes to all methods, so the current benchmarks may not hold.

Detailed benchmarks on speed and classification metrics for different methods have been provided in the documentation. Generally speaking, following conclusions can be made:

• CNN works best for near duplicates and datasets containing transformations.
• All deduplication methods fare well on datasets containing exact duplicates, but Difference hashing is the fastest.

🤝 Contribute

We welcome all kinds of contributions. See the Contribution guide for more details.

📝 Citation

Please cite Imagededup in your publications if this is useful for your research. Here is an example BibTeX entry:

@misc{idealods2019imagededup,
  title={Imagededup},
  author={Tanuj Jain and Christopher Lennan and Zubin John and Dat Tran},
  year={2019},
  howpublished={\url{https://github.com/idealo/imagededup}},
}

🏗 Maintainers

• Tanuj Jain, github: tanujjain
• Christopher Lennan, github: clennan
• Dat Tran, github: datitran

© Copyright

See LICENSE for details.