LFNet: An end-to-end lightweight model for real-time fire detection in embedded firefighting systems

This study introduces LFNet, a novel lightweight end-to-end model designed for real-time fire detection in embedded firefighting systems. LFNet addresses the challenges of detecting flames in complex scenarios, including blurred object boundaries, smoke occlusion, small targets, and cluttered backgrounds. By leveraging a star-shaped operation and Multi-Level Contextual Attention in the backbone, Partial Convolution (PConv) for feature fusion, and a weighted Bidirectional Feature Pyramid Network for detection head optimization, LFNet achieves high detection accuracy while maintaining low computational demands. The model was trained and evaluated on the LFD dataset, which combines the public DFS dataset with 6000 self-collected real fire images, enhancing robustness in diverse fire scenarios. Experimental results demonstrate that on the LFD dataset, compared with the baseline model, LFNet improves precision by 11.8%, recall by 7.6%, reduces the number of parameters by 30%, decreases the floating-point operations (GFLOPs) by 30%, and increases the frame rate by 60%. LFNet’s efficiency and accuracy make it a promising solution for early fire detection in applications such as forest fire monitoring and industrial fire safety.