🔍 A Region-Based Convolutional Neural Network Approach to Detecting Harmful Cloaked Content for Automated Content Moderation

This project aims to detect inappropriate content in images using a pre-trained Faster R-CNN model. It leverages deep learning techniques and object detection pipelines to identify explicit or sensitive elements within an image frame.

🔍 Features

• Uses Faster R-CNN for object detection
• Filter inappropriate classes
• Image annotation and visualization for results
• Purify adversarial example

🧠 Models

• Real-ESRGAN - image enhancement
• Faster R-CNN - main detection model
• ResNet50 - backbone architecture

🗂️ Datasets

To maintain class balance, 10K images were selected from each dataset (20K total). Using RoboFlow, we applied data augmentation (horizontal/vertical flips), expanding the dataset to 40K images.

• Large Scale Porngraphic Dataset: 50K annotated images
• Harmful Object Detection Dataset: 10k images

⚙️ Installation

1. Create a virtual environment

   python -m venv venv

2. Activate the virtual environment

   venv\Scripts\activate

3. Install required dependencies
   • Local (with GPU)

     pip install -r requirements.txt

4. Setup modal

   python -m modal setup

5. Run the application
   • Local

     python main.py

   • Modal

     modal serve main.py

Note: This requires ffmpeg to be installed in your local machine. You can get it here FFmpeg Download

🙌 Acknowledgements

We acknowledge the original development of Real-ESRGAN by Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. We also recognize the partial implementation and contributions provided by Igor Pavlov, Alex Wortoga, and Emily.