摘要
针对小目标检测发展缓慢、检测效果差以及公共场所吸烟防控效果差等问题,提出一种基于YOLOv5算法融合SENet注意力机制的香烟目标检测模型。该模型首先在主干网络Backbone中加入Squeeze-and-Excitation Networks(SENet)注意力机制,SENet采用了一种全新的特征重标定策略,通过自学习的方式自动获取每个特征通道的权重,增大香烟目标重要特征通道的权重值。其次使用CIoU作为定位框回归损失函数,归一化中心点距离和纵横比快速实现真实框和预测框的匹配。将YOLOv5和改进后的模型在自建数据集做对比实验,结果表明,该模型精准率达到0.85,召回率达到0.75,F1达到0.797,平均精度均值(mAP)达到0.81。与其他目标检测领域的主流模型比较,检测性能有了提升。
Aiming at the slow development of small target detection,poor detection effect and poor smoking prevention and control effect in public places,a cigarette target detection model based on YOLOv5 algorithm combined with SENet attention mechanism is proposed.Firstly,the model adds the Squeeze-and-Excitation Networks(SENet)attention mechanism to the backbone network.SENet adopts a new feature recalibration strategy to automatically obtain the weight of each feature channel through self-learning.The weight value of the important feature channel of the big cigarette target.Secondly,CIoU is used as the location box regression loss function,and the center point distance and aspect ratio are normalized to quickly match the real box and the predicted box.Comparing YOLOv5 and the improved model in the self-built data set,the results show that the accuracy of this model reaches 0.85,the recall rate reaches 0.75,the F1 reaches 0.797,and the mAP reaches 0.81.Compared with other mainstream models in the field of object detection,the detection performance has been improved.
作者
李丹妮
栾静
穆金庆
LI Dan-ni;LUAN Jing;MU Jin-qing(School of Computer Science and Technology,Xinjiang Normal University,Urumqi 830054,China)
出处
《软件导刊》
2023年第1期229-235,共7页
Software Guide
