Meme Harm Detection using YOLOv8

A computer vision project that classifies memes as harmful or safe using YOLOv8 object detection. Built and trained on Google Colab with GPU acceleration, supporting both English and Arabic meme datasets.

---

Overview

Social media platforms host millions of memes daily — some harmless, others potentially harmful to individuals or communities. This project trains YOLOv8 models to automatically detect and classify memes into distinct categories, enabling content moderation at scale.

Two separate models are trained:

• English Memes Model — general meme classification
• Arabic Memes Model — Arabic-language meme classification

---

Categories

Label	Description
normal	Neutral, harmless meme content
harmful-to-individual	Content that targets or harms individuals
community-society-organization	Safe community/social content

---

Results

Model	mAP50	mAP50-95	Precision	Recall
English Memes	0.685	0.661	0.456	0.851
Arabic Memes	0.992	0.992	0.957	0.973

The Arabic model was trained on a more balanced and curated dataset, yielding significantly higher performance.

---

Project Structure

Memes/
├── MemesData.ipynb          # Training pipeline for English memes
├── ArabicMemesData.ipynb    # Training pipeline for Arabic memes
└── README.md

---

Tech Stack

• Model: YOLOv8 nano (yolov8n.pt) — Ultralytics
• Training: Google Colab (Tesla T4 GPU, CUDA 12.2)
• Optimizer: AdamW
• Framework: Python, PyTorch

---

Training Configuration

Hyperparameter	Value
Epochs	300
Batch Size	16
Image Size	640 × 640
Optimizer	AdamW
Early Stopping Patience	200
Confidence Threshold (inference)	0.25

---

How to Run

Prerequisites

pip install ultralytics

1. Clone the repo

git clone https://github.com/Afzal632/Memes.git
cd Memes

2. Open in Google Colab

Upload either notebook to Google Colab and mount your Google Drive:

from google.colab import drive
drive.mount('/content/drive')

3. Prepare your dataset

Place your dataset in Google Drive following this structure:

/content/drive/MyDrive/
└── your-dataset/
    ├── data.yaml
    ├── train/
    │   ├── images/
    │   └── labels/
    ├── valid/
    │   ├── images/
    │   └── labels/
    └── test/
        ├── images/
        └── labels/

4. Train

Run all cells in the notebook. The pipeline will:

1. Check GPU availability
2. Install dependencies
3. Train the model (resumes from checkpoint if last.pt exists)
4. Validate on the test set
5. Run inference and display predictions

---

Training Pipeline

# Train
!yolo task=detect mode=train model=yolov8n.pt \
    data=data.yaml \
    epochs=300 \
    batch=16 \
    imgsz=640 \
    optimizer=AdamW \
    patience=200 \
    plots=True

# Validate
!yolo task=detect mode=val model=runs/detect/train/weights/best.pt data=data.yaml

# Predict
!yolo task=detect mode=predict model=runs/detect/train/weights/best.pt \
    conf=0.25 source=datasets/test/images

---

Notes

• The English model has high recall (0.851) but lower precision (0.456), meaning it catches most harmful memes but produces some false positives.
• The Arabic model achieves near-perfect scores, though results should be interpreted carefully given the smaller test set size.
• Models are saved as best.pt and last.pt and can be reused for inference without retraining.

---

Use Cases

• Social media content moderation
• Automated flagging of harmful meme content
• Research on Arabic and English online harmful content
• Dataset labeling assistance

---

License

This project is intended for academic and research purposes.