DG-YOLO: A Novel Efficient Early Fire Detection Algorithm Under Complex Scenarios

In reality, it is important to control fires in their early stages. However, the early stages of a fire are characterized by small flames with blurred edges. Additionally, the interference in complex scenarios involving occlusion, light interference, and fire-like objects leads to a high leakage rate and false detection rate of existing target detection methods in early fire detection. To address the above problems, this paper proposes a novel and efficient method for early fire detection in complex scenarios, called DG-YOLO. Firstly, a deformable attention (DA) is introduced in the YOLOv8 backbone. Focusing on small fire features, it enhances the anti-interference ability of the model in complex scenes. Secondly, the addition of a lightweight feature extraction module (GSC2f) gives the model a rich gradient flow to capture early flame edge features, thus enabling effective multi-scale feature fusion. Finally, to address the limitations of small early flames and blurred edges, we introduce a small-target detector. It effectively captures the shape and texture information of early fires in complex scenes and reduces the leakage rate and false alarm rate. Comprehensive experiments have been conducted on a dataset of real-life scenarios. The results of the study show that the F1 score and mAP50 metrics are improved by an astonishing 9.77% and 10.7%, respectively. The leakage rate and false alarm rate are effectively reduced. Meanwhile, comparison experiments show that DG-YOLO surpasses the current advanced technology. The efficiency of the model for early fire detection in complex scenarios is demonstrated.