基于频空卷积改进YOLOv11的火灾检测算法

Fire Detection Algorithm Based on Improved YOLOv11 with Frequency-Spatial Convolution

摘要: 本文针对火灾检测任务提出了一种基于频空卷积(Frequency-Spatial Convolution, FSConv)改进的火灾检测算法——FSConv-YOLOv11。通过引入FSConv模块重构了YOLOv11的主干与颈部特征提取网络，实现了检测精度与计算效率的双重优化。具体而言，FSConv模块利用哈尔离散小波变换(Haar DWT)将特征图无损解耦为高频细节与低频背景，利用高频特征引导空间注意力机制，实现了频域与空域特征的深度融合；同时，该模块采用无参数的小波变换与深度可分离卷积，大幅削减了传统标准卷积带来的计算冗余。实验结果表明，改进后的FSConv-YOLOv11算法检测精度(mAP@0.5)达到了0.632，较YOLOv11提升了1.2%。另外，本文所提算法的复杂度大大降低，具体包括：模型参数量减少了15.4%，梯度计算量降低了15.4%，GFLOPs降低了9.2%，本文所提算法为复杂场景下的火灾智能预警及边缘端部署提供了高效可靠的技术支撑。

Abstract: Aiming at the fire detection task, this paper proposes an improved fire detection algorithm named FSConv-YOLOv11 based on Frequency-Spatial Convolution (FSConv). The FSConv module is embedded to restructure the backbone and neck feature extraction networks of YOLOv11, realizing dual optimization of detection accuracy and computational efficiency. Specifically, the FSConv module adopts the discrete Haar Wavelet Transform (Haar DWT) to losslessly decouple feature maps into high-frequency detail features and low-frequency background features, and high-frequency features are used to guide the spatial attention mechanism to achieve deep fusion of frequency-domain and spatial-domain features. Meanwhile, the module combines parameter-free wavelet transform with depthwise separable convolution to greatly eliminate computational redundancy caused by traditional standard convolution. Experimental results verify that the improved FSConv-YOLOv11 obtains an mAP@0.5 of 0.632, which is 1.2% higher than the original YOLOv11. Besides, the proposed algorithm achieves obvious reduction in model complexity: the number of model parameters is reduced by 15.4%, the amount of gradient computation drops by 15.4%, and GFLOPs is decreased by 9.2%. The proposed algorithm provides efficient and reliable technical support for intelligent fire early warning and edge-end deployment in complex scenarios.