Skip to content

Neural network for creating distortion while keeping embeddings as close as possible

License

Notifications You must be signed in to change notification settings

ZamDimon/distortion-generator

Repository files navigation

🖼️ ⁉️ Distortion Generator

| [Journal Paper] | [Conference Paper] | [Papers with Code] | [Citation] |

Neural network for creating distortion while keeping embeddings as close as possible. Part of the research paper "Unrecognizable Yet Identifiable: Image Distortion with Preserved Embeddings."

The code is written in TensorFlow v2.12.

Example Generations

ℹ️ Paper info

Unrecognizable Yet Identifiable: Image Distortion with Preserved Embeddings

Dmytro Zakharov1, Oleksandr Kuznetsov1,2, Emanuele Frontoni2

1 V. N. Karazin Kharkiv National University, Ukraine

2 University of Macerata, Italy

Preprint at arXiv

📁 Structure

The project is structured as follows:

File/Folder Description
cli.py CLI for running training or evaluation
src All source files for training and evaluating the models
images Images with example generations, evaluation plots etc.
hyperparams_embedding.json Hyperparameters for training the embedding model
hyperparams_generator.json Hyperparameters for training the generator model
dataset Dataset which was used for training (actually, the portion of it since we do not want to put everything into the repository)
models Models' weights after the training. Just so you know, the newest versions of the generator are not included since they weigh too much for GitHub to handle.

Citation

@misc{zakharov2024unrecognizable,
      title={Unrecognizable Yet Identifiable: Image Distortion with Preserved Embeddings}, 
      author={Dmytro Zakharov and Oleksandr Kuznetsov and Emanuele Frontoni},
      year={2024},
      eprint={2401.15048},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}