Explicit Content Detection using Vision Transformers

Overview

This repository focuses on detecting explicit content in images using Vision Transformers (ViTs) and Swin Transformers. The project explores multiple architectures, including convolution-based patch extraction and autoencoder-based feature extraction, to improve detection accuracy.

Models Implemented

1. Feed Forward Patch Extraction - Standard ViT approach with feed-forward networks.
2. Convolution-Based Patch Extraction - Using CNNs to extract patches before feeding them into transformers.
3. Swin Transformer Model - A hierarchical vision transformer achieving 98.5% accuracy.
4. Autoencoder Feature Extraction (Upcoming) - Enhancing feature representation with autoencoders.

How to Use This Repository

Clone the repository using:

  git clone https://github.com/yaseeng-md/Explicit-Content-Detection
  cd Explicit-Content-Detection

Install the necessary dependencies from the requirements.txt file.

pip install -r requirements.txt

Run the provided Jupyter notebooks to preprocess the data, train models, and make predictions.

Prerequisits

Before you continue with the implemntation, Take the help of Dataset Collection.py and Remove Corrupt Files.py in Helpers.

Remove Corrupt Files.py Helps you to remove corrupted files from the exitsing dataset.

Dataset Collection.py You can decide the amount of dataset that you want to train and experiment on.

Trained Models

Download the trained Models from here !

Dataset

Provided on request.

Results

Model	Traning Accuracy	Training Loss	Validation Accuracy	Validation Loss
ViT (Feed Forward)	78.95%	0.4638	81.89%	0.4351
ViT (CNN Patch Extractor)	88.05%	0.430	81.86%	0.4893
Swin Transformer	98.35%	0.0458	86.89%	0.3443

Future Work

• Implement Feed Forward Autoencoder-based Feature Extraction.
• Implement Convlution Autoencoder-based Feature Extraction.