基于YOLOv11的火灾检测算法研究

Research on Fire Detection Algorithm Based on YOLOv11

摘要: 由于火灾事故的频繁发生，火灾检测技术在保障生命财产安全和推动社会可持续发展方面具有重要意义。近年来，基于计算机视觉和深度学习的火灾检测方法因其高效性和较强的适应能力，逐渐成为火灾检测领域的研究热点。鉴于现有火灾检测算法难以兼顾多尺度目标与小目标检测，本文基于YOLOv11提出一种改进的火灾检测模型YOLOv11_VAE。首先，为弥补现有数据集的不足，本文构建了一个包含7687张图像、专注于高危场景下初期微小火焰的专用数据集。其次，为提升对微小火焰的感知能力，设计了多尺度卷积与注意力增强模块(MCAE)，通过“生成–选择”策略捕捉并增强不同尺度的关键火焰特征。为降低复杂背景导致的误报，提出了全局–局部特征聚合模块(GLFA)，通过融合场景级上下文信息提升模型的判别能力。在自建数据集上的大量实验证明了所提方法的有效性。结果表明，YOLOv11_VAE的mAP@0.5值达到0.964，相较于YOLOv11n基线模型提升了3.3个百分点，同时模型参数量仅为2.70 M，计算量为6.6 GFLOPs，实现了高精度与高效率的平衡。YOLOv11_VAE为高危环境下的早期火灾预警提供了一种有效且高效的解决方案。

Abstract: Due to the frequent occurrence of fire accidents, fire detection technology is of great significance in ensuring the safety of life and property and promoting the sustainable development of society. In recent years, fire detection methods based on computer vision and deep learning have gradually become a research hotspot in the field of fire detection due to their high efficiency and strong adaptability. In view of the fact that the existing fire detection algorithms are difficult to take into account the detection of multi-scale objects and small objects at the same time, this paper proposes an improved fire detection model YOLOv11_VAE based on YOLOv11. Firstly, in order to make up for the shortcomings of the existing data sets, this paper constructs a dedicated data set containing 7687 images focusing on the early tiny flames in high-risk scenes. Secondly, in order to improve the perception ability of tiny flames, a multi-scale convolution and Attention Enhancement module (MCAE) is designed to capture and enhance key flame features at different scales through the “generation-selection” strategy. In order to reduce false positives caused by complex backgrounds, a Global-local feature aggregation module (GLFA) was proposed to improve the discriminative ability of the model by fusing scene-level context information. Extensive experiments on self-built datasets demonstrate the effectiveness of the proposed method. The results show that the mAP@0.5 value of YOLOv11_VAE reaches 0.964, which is 3.3 percentage points higher than that of the YOLOv11n baseline model. At the same time, the model parameter number is only 2.70 M, and the calculation amount is 6.6 GFLOPs, achieving a balance between high accuracy and high efficiency. YOLOv11_VAE provides an effective and efficient solution for early fire warning in high-risk environments.