🛡️ Shielding Social Media: Detection of Toxic Memes for Automated Moderation

📌 Abstract

This project tackles the challenge of moderating toxic content in internet memes, which combine images and text to spread potentially harmful or offensive messages. We propose a multimodal detection system that uses OCR, NLP, and Vision-Language Fusion via SigLIP to classify memes as toxic or non-toxic, ensuring safer social media environments.

---

🧠 System Architecture

Architecture Diagram: Fusion-based Toxic Meme Classifier with PaddleOCR, Kosmos-2, Vision Transformer, and SigLIP.

🔄 Flow Description:

1. Input Meme: Raw image meme with visual and textual content.

2. Text Preprocessing:

   • Uses PaddleOCR and KOSMOS-2 for text extraction.

3. Image Preprocessing:

   • Image resized and transformed into pixel tensors.

4. Feature Embedding:

   • Text and image inputs are separately embedded.
   • Passed into SigLIP model for multimodal fusion.

5. Classification:

   • Uses Sigmoid activation + Cross Entropy Loss.
   • Optimized to predict Toxic or Non-Toxic.

---

🚀 Features

• 🔍 Detects toxicity in memes using a deep learning fusion approach.
• 🔤 Supports OCR from meme text using PaddleOCR & KOSMOS-2.
• 👁️‍🗨️ Uses SigLIP (Google) for image-text fusion and classification.
• 📈 Provides performance metrics and visualizations.

---

🛠️ Tech Stack

Component	Technology
Language	Python 3.8+
OCR Engine	PaddleOCR, KOSMOS-2
Text Encoder	BERT (transformers)
Image Encoder	Vision Transformer (ViT)
Fusion Model	SigLIP
DL Framework	PyTorch
Visualization	Matplotlib, Seaborn

🔮 Future Scope

🌍 Multilingual toxic meme detection

🎥 Video meme frame-based detection

🌐 Web portal for real-time moderation

📦 Deploy as browser extension / REST API

📄 License

This project is licensed under the MIT License. See the LICENSE file for more details.

⭐ Support

If you found this project useful, please consider giving it a ⭐ and sharing it with others!