摘要：

在林火检测领域中, 由于存在背景环境复杂、火焰初期目标过小、火焰形状及分布不规则等因素, 且对检测模型的部署要求较高, 现有的基于深度学习的目标检测方法难以发挥很好的效果。为更好地预防森林火灾的发生及扩散, 保护森林环境及人民的生命财产安全, 提出一种基于RF-YOLOv8s的轻量化林火实时检测算法。该方法基于YOLOv8模型, 首先引入一种基于感受野注意力机制的卷积模块RFCBAMConv, 通过强调感受野的空间特征以强化对火焰特征的学习能力。其次在特征融合部分采用轻量化的跨尺度特征融合模块(CCFM), 在减少模型参数和计算量的同时增强模型对尺度变化的适应性和对小尺度对象的检测能力。同时采用动态检测头(DyHead), 借助注意力机制统一并强化多尺度信息, 提高模型的检测效果。此外, 使用Inner-CIoU作为新的边界框损失函数, 通过控制辅助边框来克服边框回归的泛化性不足并加快模型的收敛速度。实验结果表明, 在林火图像数据上, RF-YOLOv8s林火检测算法的平均精度为90.2%, 模型参数量仅8.88×10⁶, 较YOLOv8算法平均精度提升了2.5百分点的同时减少了20.2%的模型参数量, 在具有更强的检测能力的同时, 满足了林火检测领域对模型轻量化的要求, 确保了该算法在森林火灾检测场景上的实用性。

关键词: 森林火灾检测, 感受野注意力, 特征融合, YOLOv8算法, 损失函数

Abstract:

In the field of forest fire detection, existing target detection methods based on deep learning find it difficult to achieve good results because of the complex background environment, small initial targets of the flame, irregular flame shape and distribution, and complex requirements for the deployment of a detection model. To better prevent the occurrence and spread of forest fires, protect the forest environment, and ensure the safety of people and property, a lightweight real-time forest fire detection method called RF-YOLOv8s is proposed. Based on the YOLOv8 model, this method first introduces a convolution module RFCBAMConv based on receptive field attention mechanism, which emphasizes the spatial characteristics of the receptive field to strengthen the learning ability of flame characteristics. Second, in the feature fusion part, a lightweight Cross-Scale Feature Fusion Module (CCFM) is used to reduce model parameters and computation while enhancing the model's adaptability to scale changes and detection ability for small-scale objects. Simultaneously, a Dynamic detection Head (DyHead) is used to unify and strengthen multi-scale information with the help of an attention mechanism, improving the model's detection effect. In addition, Inner-CIoU is used as a new bounding box loss function to overcome the generalization deficit of bounding box regression and accelerate the model's convergence speed by controlling the auxiliary bounding box. The experimental results show that on forest fire image data, the average accuracy of RF-YOLOv8s forest fire detection algorithm is 90.2% and the number of model parameters is only 8.88×10⁶; these are 2.5 percentage points higher than the average accuracy and 20.2% smaller than the number of model parameters of the YOLOv8 algorithm. While having stronger detection ability, it meets the requirements of lightweight models in the field of forest fire detection, ensuring the algorithm's practicability in forest fire detection scenarios.

Key words: forest fire detection, reception filed attention, feature fusion, YOLOv8 algorithm, loss function