Trypophobia images detector based on deep neural networks and utilities
Ever wanted to censor images on the web using deep neural networks?
A deep learning project by Artur Puzio and Grzegorz Uriasz made as part of an internship at deepsense.ai sponsored by The Polish Children's Fund and supervised by Piotr Migdał.
A browser extension is available for Mozilla Firefox. Try it here.
The utilities contained in the utils
folder are small programs and scripts useful in generating the data set and easing the usage of the deep learning lab Neptune.
Note: The provided images may be or not be subject to copyright. By downloading the dataset you agree to use it only for research purposes.
Images have been divided into 4 folders
/valid/trypo
- 500 random trypophobia triggering images/valid/norm
- 500 random neutral images/train/trypo
- rest of the trypophobia triggering images/train/norm
- rest of the neutral imagesAnyone interested in the "raw" unprocessed data please send us an email.
The models were made in the Keras machine learning framework and are compatible with the Keras.JS javascript library. The models were trained on the Google Computing Platform using Neptune. This repository contains some of the models together with the training results. We examined the performance of different size models and decided to aim for one with less than 20k parameters. We achieved up to 90% accuracy and 0.27 log-loss on the validation set. Additionally, some models with <10k parameters came close to achieving these results.
The browser plugin censors images encountered while browsing the web. It uses a supplied trained model to determine which images are safe to reveal and which a warning must be issued for. The extension is a WebExtension and was tested on Mozilla Firefox. Currently the extension works on most sites. You can try it here.