Abstract: Aiming at the lack of detection accuracy and speed bottleneck of traditional fire detection algorithms, especially for the recognition of small-scale incipient fires and large-scale rapidly spreading fires, this paper proposes a lightweight fire detection algorithm based on multi-scale feature fusion. Firstly, the EDBAN module is designed to replace the C2f module in YOLOv8, to improve the generalization ability and adaptability of the model, especially the accuracy when dealing with multi-scale fire scenarios. Secondly, the original BiFPN structure is improved to adapt to the YOLOv8 model structure, and the Weighted Blend module is designed to weight and fuse the features at each layer, which enhances the feature characterization ability and reduces the risk of missed detection. Finally, the LOTT detection module is further proposed to replace the traditional YOLOv8 detection, which achieves the accuracy and stability of the detection performance while maintaining the lightweight through a series of group convolution and scale adjustment operations. Through experiments on scene-rich fire datasets, the results show that the improved YOLOv8 algorithm reduces the parameter amount by 58.3% and the computation amount by 34.5% based on the baseline model, while the mAP improves by 2.6 percentage points, which basically meets the demand of real-time fire detection.

Key words: YOLOv8, lightweighting, fire detection, target detection, BiFPN

摘要： 针对传统火灾检测算法存在的检测精度不足及速度瓶颈，特别是对于小规模初发火情与大规模迅速蔓延火灾的识别难题，研究提出一种基于多尺度特征融合的轻量级火灾检测算法，设计了EDBAN模块以替代YOLOv8中的C2f模块，提升模型的泛化能力和适应性，尤其是在处理多尺度火灾场景时的精准度。改进原有的BiFPN结构适配YOLOv8模型结构，并设计Weighted Blend模块对各层特征进行加权融合，增强特征的表征能力，降低漏检风险。进一步提出LOTT检测模块，以替代传统的YOLOv8检测，通过一系列组卷积和尺度调整操作，实现了在轻量化的同时保持了检测性能的准确性和稳定性。通过在场景丰富的火灾数据集上进行实验，结果表明，改进的YOLOv8算法在基准模型的基础上参数量减少了58.3%、计算量减少了 34.5%，同时 mAP 提升了2.6个百分点，基本满足火灾实时检测的需求。

关键词: YOLOv8, 轻量化, 火灾检测, 目标检测, 加权双向特征金字塔（BiFPN）