To solve the problem of high false detection rates in tunnel fire detection caused by the complexity of tunnel environments based on the YOLOv8n network model, an improved tunnel fire detection algorithm was proposed.First, in the backbone network, the FasterNet network was used for replacement while retaining the original SPPF module to achieve more comprehensive feature extraction; Secondly, in order to improve the detection accuracy of the model for irregular targets in the complex background, the D-LKA attention mechanism was introduced in the C2f module; Finally, Focaler-IoU to optimize the model loss function was introduced, which further reducing the problem of false positives or false negatives caused by distractors. The experimental results show that compared with YOLOv5, YOLOv7 and the original models of YOLOv8n, the accuracy of the improved model is increased by 7.6%, 5.6%, and 3.5% respectively, and the average accuracy means are increased by 8.3%, 7.7%, and 5.1% respectively. Compared with other YOLOv8n-based improved algorithms, the mean average precision of our proposed model is increased by 3.3% and 6.4% respectively.
FasterNet Block模块是FasterNet网络模型中的主要特征提取模块,其中的PConv仅对一部分通道应用常规卷积以进行特征的提取,而其他通道保持不变。由于通道数的减少,模型参数量随之降低,特征提取过程中的计算量也得以减少,Conv、PConv的浮点运算数(floating point operations per second,FLOPs)如式(1)所示。
可变形大核注意力[12](deformable large kernel attention,D-LKA)最早作为一种用于医学图像分割的有效注意力机制而被提出,采用大卷积核并结合可变形卷积,有效提升了模型对小目标和不规则形状目标的检测能力,尤其对复杂背景或不同光照条件下的目标检测具有突出效果。因此,本文将D-LKA机制融入C2f模块[13],进一步强化了在复杂或昏暗背景下模型对火焰及烟雾的检测能力,性能有所提升。
ChenT H, WuP H, ChiouY C.An early fire-detection method based on image processing[C]//2004 International Conference on Image Processing,Singapore:IEEE,2004:1707-1710.
[2]
CelikT, DemirelH, OzkaramanliH,et al.Fire detection using statistical color model in video sequences[J].Journal of Visual Communication and Image Representation,2007,18(2):176-185.
CaoY C, YangF, TangQ F,et al.An attention enhanced bidirectional LSTM for early forest fire smoke recognition[J].IEEE Access,2019,7:154732-154742.
[5]
BarmpoutisP, DimitropoulosK, KazaK,et al.Fire detection from images using faster R-CNN and multidimensional texture analysis[C]//ICASSP 2019-2019 IEEE International Conference on Acoustics,Speech and Signal Processing.Brighton:IEEE,2019:8301-8305.
AzadR, NiggemeierL, HüttemannM,et al.Beyond self-attention:deformable large kernel attention for medical image segmentation[C]//2024 IEEE/CVF Winter Conference on Applications of Computer Vision.Waikoloa:IEEE,2024:1276-1286.