Federated Learning for Smoke and Fire Detection Models Optimization

Incidents of fire and smoke can lead to a lot of damage and may cause some casualties. Modern fire and smoke detection approaches lack generalization as they are only trained to perform well on a specific type of data and scenes. This research paper explores different deep-learning and federated learning techniques for detecting fire and smoke in various scenarios due to the significant safety risks they pose. We present two different approaches which are centralized training and federated training with multiple clients and datasets from different scenes like indoor fire and outdoor fire scenes. Additionally, we performed multiple quantization techniques for efficient CPU utilization and model size reduction so that our model can work in real-time on embedded systems. An automated fire and smoke detection early alarm system is employed to alert administrators or security personnel. We also perform distance estimation to determine the distance of the fire from the camera. The results demonstrate that centralized training achieves an impressive accuracy of $77 \%$. The proposed federated models showed varying results for different types of data. Moreover, the averaged federated learning model gave good accuracy in both outdoor and indoor scenes. Our proposed approach paves the way for the use of federated learning for early fire and smoke detection systems in multiple environments simultaneously.