This project aims to detect protest activity in images using deep learning techniques. Inspired by the paper "Protest activity detection and perceived violence estimation from social media images", the model classifies whether an image represents a protest or not.
With the rise of social media as a real-time news source, being able to detect protest-related content has applications in:
- Public safety
- Sociopolitical research
- Crisis response automation
This project explores how EfficientNet, advanced data augmentation, and comparative model analysis can improve image-based protest detection performance.
- β Binary image classification: Protest vs Non-Protest
- β Backbone: EfficientNet-B0 to B3
- β
Data augmentation using
Albumentations
for robustness - β Comparative analysis with other CNNs (e.g., ResNet, VGG)
- β Evaluation with accuracy, ROC-AUC, confusion matrix, etc.
This project uses the dataset provided in the original paper:
Wang, Z., Hale, S. A., Adelani, D. I., & Hanna, A. (2018). Protest activity detection and perceived violence estimation from social media images.
You can refer to the paper for data access.
- Protest Activity Detection Paper (Wang et al., 2018)
- EfficientNet: Rethinking Model Scaling
- Albumentations Library
PRs and feedback are welcome! If youβd like to contribute, fork this repo and submit a pull request.