A Fire and Smoke Detection Model Based on YOLOv8 Improvement

Abstract

The warning of fire and smoke provides security for people's lives and properties. The utilization of deep learning for fire and smoke warning has been an active area of research, especially the use of target detection algorithms has achieved significant results. For improving the fire and smoke detection performance of model in different scenarios, a high-precision and lightweight improvement based on the model of You Only Look Once (YOLO), is developed. It utilizes partial convolutions to reduce the complexity of model, and add an attention block to acquire the cross-space learning capability. In addition, the neck network is redesigned to realize bidirectional feature fusion. Experiments show that it has significantly improved the results for all metrics in the public Fire-Smoke dataset, and the size of the model has also been widely reduced. Comparisons with other popular target detection models under the same conditions indicate that the improved model has the best performance as well. In order to have a more visual comparison with the detectability of the original model, the heatmap experiments are also established, which also demonstrate that it is characterized by less leakage rate and more focused attention.