Fire detection algorithm with multi-scale feature focus and diffusion

Flame and smoke detection in fire detection systems can suffer from low accuracy and high false detection rates due to background and environmental factors. To address this, a multi-scale feature focusing and diffusion multilayer fusion fire detection algorithm is proposed, using YOLOv8s as the baseline model. This approach introduces large separable convolutional kernel attention at the end of the backbone network to expand the receptive field, enhance the understanding of spatial feature relationships, and improve feature extraction capabilities. The neck network is enhanced by refining the PANet structure, adding a feature focusing module for multi-scale feature fusion, and incorporating a feature diffusion mechanism to boost the model's utilization of global context information and multi-layer feature integration. The detection head is upgraded with a dynamic detection head featuring unified attention, which consolidates three types of attention to improve detection accuracy. After algorithm optimization and testing, the improved model shows a 1.4 percentage point increase in [email protected], a 2.6 percentage point increase in [email protected]:0.95, and a detection speed of 84 fps, outperforming other mainstream object detection algorithms.