YOLO-DKM: A Flame and Spark Detection Algorithm Based on Deep Learning

In the field of fire prevention and industrial safety monitoring, it is very important to accurately and efficiently detect fires and deal with some factors that may affect fires in the early stages. In the field of computer vision, existing fire detection algorithms often have problems such as low detection accuracy and algorithm error detection. Moreover, for some fire warning factors, the current focus is mainly on detecting smoke, while existing detection algorithms are still relatively lacking in detecting other warning factors. In order to explore other fire warning factors and enhance the detection capability of fires, this paper proposes an improved YOLOv8 fire and spark detection algorithm-YOLO-DKM. Improvements have been made to the backbone network and neck network. Firstly, the SKAttention attention mechanism has been introduced into the backbone network, which can adaptively adjust the attention weights at different scales, effectively improving the detection of small flame targets or tiny granular sparks. And integrate DSConv into the C2f module, relying on dynamic characteristics to adaptively adjust convolution operations according to different scene features for more flexible local region feature extraction, capturing local features of flames and sparks. Introducing CBAM attention mechanism into the neck network can help reduce background interference, enhance model perception, and improve detection accuracy and recall. The experimental results show that compared with the original YOLOv8 algorithm, the YOLO-DKM algorithm improves accuracy by 5.4% and recall by 3.6%, proving the effectiveness of the improved algorithm in detecting flames and sparks.