TinyEmergencyNet

TinyEmergencyNet investigates efficient aerial-scene classification for emergency-response applications, where unmanned aerial platforms must operate under strict limits on memory, energy, and compute.

Motivation

Conventional convolutional neural networks can provide strong image-classification accuracy, but their parameter counts and floating-point workloads make them difficult to deploy on constrained airborne and edge platforms. The project asks how much of that computational cost can be removed while retaining a useful model for real-time situational awareness.

Result

The resulting network is designed around a small computational and memory footprint, with an architecture intended for practical embedded and FPGA deployment. The project became the central model-design contribution of my MSc research and was published in the Journal of Real-Time Image Processing in 2024.