Abstract:

【Objective】Forest fires pose a significant threat to the natural environment and human safety, thus timely and accurate detection of fire sources is demanding. However, the complex forest environment characterized by high tree density, ground litter accumulation, and dense canopies creates substantial challenges for effective fire detection. To address these issues, this study proposes a novel forest fire detection method, SBP-YOLOv7.【Method】The proposed method incorporated three key innovations. First, an attention mechanism-enhanced downsampling module (BRA-MP) was introduced to improve feature recognition during downsampling, enhancing the model’s ability to detect small targets by boosting feature representation and semantic relevance. Second, the extended partial convolution efficient layer aggregation module (EP-ELAN) was integrated into the model’s backbone, effectively reducing redundant computations and minimizing model parameters. Finally, a Slim-neck neck module was employed for feature fusion, ensuring high accuracy while lowering computational costs. 【Result】Comparative evaluations on a forest fire dataset demonstrate that the SBP-YOLOv7 model achieves an AP score of 87.0%, representing 2.3% improvement over the original YOLOv7. Additionally, the model reduces parameter count by 22.77% and computational cost by 17.13%.【Conclusion】Compared with the traditional YOLOv7 algorithm, the proposed SBP-YOLOv7 model offers superior accuracy and efficiency, enabling faster and more precise detection of forest fires even in challenging environments.

Key words: deep learning, forest fire detection, forest scenes, YOLOv7, attention mechanism