摘要
火灾是一种极具破坏性的灾害,对火焰和烟雾的检测有助于及时发现火灾,以便采取有效措施将损失最小化.现有的算法对火焰烟雾的检测精度较低,不能做到精确识别小目标火焰和烟雾.为了进一步提高检测准确率,提出了基于YOLOv5s算法进行改进的火焰烟雾检测算法FS-YOLO.首先,在主干网络的C3模块中融合CA注意力机制来增强模型对图像特征的感知能力;其次,为了实现高效多尺度特征融合,将双向加权特征金字塔网络(BiFPN)用于颈部的多尺度信息融合;此外,在主干网络中加入由混合卷积层和普通卷积层组成的CSPCM模块,以便通过少量计算代价来提取火焰烟雾显著特征;最后,为提高边界框回归准确率,采用了基于最小点的MPDIoU损失函数和ReLU激活函数加速网络的训练和推理.在Fire and Smoke数据集上的实验结果表明:FS-YOLO达到了mAP@0.5上0.606和mAP@0.5-0.95上0.275的检测精度,相较于YOLOv5s分别提升5.21%和8.27%,FS-YOLO在实际运行中的速度为303 FPS,实现了高精度快速的实时火焰烟雾检测.
Fire is a highly destructive disaster,the detection of fire and smoke helps to distinguish fire promptly so that effective measures can be taken to minimize losses.The existed algorithms had low detection accuracy for fire-smoke,which could not accurately recognize the fire and smoke in the real world.In oder to further improve the accuracy of detection,an improved YOLOv5s fire smoke detection algorithm was proposed,named FS-YOLO.Specifically,CA attention mechanism was embedded into C3 blocks to facilitate the feature perception ability in the backbone.Then,in order to realize enhanced the multi-scale feature fusion,bidirectional weighted feature pyramid network named BiFPN,was employed in the neck to realize cross-layers multiscale information fusion.Besides,the CSPCM module which consisted of a hybrid convolutional layer and a vanilla layer were inserted into the backbone network with a fraction of computational cost to acquire discriminative fire and smoke features.Finally,a mini-distance computation based MPDIoU bounding box loss function was utilized to boost detection accuracy and the ReLU activation function was applied to speed up the model training and inference process.The experimental results on the Fire and Smoke dataset showed that FS-YOLO achieved a detection accuracy of 0.606 mAP@0.5 and 0.275 mAP@0.5-0.95,which was 5.21%and 8.27%higher than the baseline YOLOv5s,respectively.The inference speed of FS-YOLO in the real application was at 303 FPS,realizing the real-time fire and smoke detection with high accuracy and speed.
作者
徐慧英
胥玲玲
李毅
朱信忠
陶珏强
黄晓
XU Huiying;XU Lingling;LI Yi;ZHU Xinzhong;TAO Jueqiang;HUANG Xiao(School of Computer Science and Technology,Zhejiang Normal University,Jinhua 321004,China;College of Engineering,Zhejiang Normal University,Jinhua 321004,China;College of Education,Zhejiang Normal University,Jinhua 321004,China)
出处
《浙江师范大学学报(自然科学版)》
2025年第3期267-276,共10页
Journal of Zhejiang Normal University:Natural Sciences
基金
国家自然科学基金资助项目(62376252,61976196)
浙江省自然科学基金重点资助项目(LZ22F030003)
国家级大学生创新创业训练计划项目创新训练重点资助项目(202410345053)。
