基于空–频域联合表征的YOLOv11改进模型及初期火灾烟雾检测应用

YOLOv11 Improved Model Based on Joint Spatial-Frequency Domain Representation and Its Application in Early Fire Smoke Detection

摘要: 针对复杂环境下初期火灾烟雾小目标、动态性检测难题，提出一种基于改进YOLOv11的高精度轻量化火灾烟雾检测算法。将空–频域聚合Mamba引入YOLOv11架构，构建MixFreSBlock模块替代原C3K2模块：通过MSPLCK模块在空域构建多尺度大感受野，同时利用MFCA模块在频域实现噪声压缩与纹理强化，二者拼接形成互补的空–频联合表征；采用SS2D模块沿四维扫描路径对联合特征进行线性复杂度全局建模，突破卷积局部性限制；最后嵌入坐标注意力机制，沿水平–垂直方向重校准特征，精准定位火焰、烟雾细长结构。实验结果表明，该模型在自建与开源数据集上mAP50达94.7%，较原YOLOv11提升0.9%，参数量仅3.1 M，计算量仅8.0 G，在精度与轻量化之间实现更优平衡。本研究为实时火灾预警系统提供了高性能、低计算成本的解决方案，可广泛应用于安防监控、智慧城市等多场景，具有重要的实际应用价值。

Abstract: To address the challenges of detecting small and dynamic early fire smoke targets in complex environments, a high-precision lightweight fire smoke detection algorithm based on an improved YOLOv11 is proposed. The spatial-frequency domain aggregation Mamba is introduced into the YOLOv11 architecture, and the MixFreSBlock module is designed to replace the original C3K2 module: the MSPLCK module constructs a multi-scale large receptive field in the spatial domain, while the MFCA module achieves noise compression and texture enhancement in the frequency domain, and the two are concatenated to form complementary spatial-frequency joint representations; the SS2D module performs linear complexity global modelling of joint features along a four-dimensional scanning path, overcoming the locality limitation of convolution; finally, a coordinate attention mechanism is embedded to recalibrate features along horizontal and vertical directions, accurately locating flame and slender smoke structures. Experimental results show that the model achieves an mAP50 of 94.7% on both custom and open-source datasets, an increase of 0.9% compared to the original YOLOv11, with only 3.1 M parameters and 8.0 G computations, achieving a better balance between accuracy and lightweight design. This study provides a high-performance, low-computation solution for real-time fire warning systems and can be widely applied in security monitoring, smart cities, and other scenarios, with significant practical value.