基于改进YOLOv5s的输电通道下的烟雾识别

Smoke Detection based on Improved YOLOv5s under Transmission Channel

摘要:

针对输电通道下出现火灾险情而难以及时发现的问题，能够在火灾初期发现形状不规则且稀薄的烟雾的产生,对于险情的控制具有重要作用。为解决此问题，提出了改进YOLOv5s网络的烟雾识别算法。该方法通过在YOLOv5s模型中引入卷积注意力模块(CBAM)，提高了对外轮廓并不明显的烟雾的特征提取能力；同时引入CARAFE特征上采样算法，扩大感知域，利用到图片中的其他信息帮助捕捉烟雾的深层特征；为捕捉到图像中目标较小的烟雾形态，利用FReLU替换原有激活函数SiLU，用二维漏斗激活函数,在引入少量计算和过拟合风险的情况下来对网络空间中的不敏感信息进行激活,进而改善视觉任务。实验结果表明，该项目改进后的检测效果相对于原始YOLOv5s网络中的查准率提高了6.8%，查全率提高了2.8%，平均精度均值提高了2.3%，检测精度提升较为明显，更有利于应用于实际情况下的烟雾检测。

Abstract:

In response to the challenge of detecting fire hazards in power transmission corridors in a timely manner, especially in the early stages of a fire when irregular and thin smoke is difficult to detect, an improved smoke recognition algorithm for YOLOv5s network is proposed. This method enhances the capability to extract features of smoke with less distinct outlines by introducing a Convolutional Block Attention Module (CBAM) into the YOLOv5s model. Additionally, it incorporates the CARAFE feature upsampling algorithm to expand the perception field and leverage other image information for capturing deep smoke features. To better detect smaller smoke patterns in the images, the SiLU activation function is replaced with FReLU, a two-dimensional funnel-shaped activation function. This modification activates insensitive information in the network space while introducing minimal computational overhead and overfitting risks, thereby enhancing visual task performance. Experimental results demonstrate that the improved algorithm in this project exhibits a 6.8% increase in precision, a 2.8% increase in recall, and a 2.3% improvement in mean Average Precision relative to the original YOLOv5s network. This significant enhancement in detection accuracy makes it more suitable for practical smoke detection applications.