VIO-CW: Mobile-Based Violence Detection

VIO-CW is a research-driven project focused on solving the challenge of running heavy deep learning models on resource-constrained mobile devices for content moderation.

The system employs Knowledge Distillation, where a complex 'Teacher' model (EfficientNet) guides a lightweight 'Student' model (MobileNetV3) to achieve high-tier accuracy with a fraction of the computational cost.

• Designed a hybrid architecture for real-time violence detection in images and video frames.
• Reduced model size significantly while maintaining a high F1-score through distillation techniques.
• Deployed a robust backend on Ubuntu VPS with Nginx and PM2 to handle model metadata and API requests.
• Integrated the model into a mobile-first environment for proactive content filtering.

Model Architecture & Optimization

Comparative analysis between teacher (EfficientNet) and student (MobileNetV3) models using Knowledge Distillation.

Mobile Integration

Seamless inference on mobile devices with optimized latency and high accuracy for real-time moderation.