基金Special Project of Innovation and Development of Shihezi University (CXFZ202103)Natural Science Foundation of Guangdong Province(2022B1515120059)+4 种基金Natural Science Foundation of Guangdong Province (2023A1515011230)National Natural Science Foundation of China(61871475,62373390)Innovation Team Project of Universities in Guangdong Province(2021KCXTD019)Meizhou City S&T Planed Projects(2021A0305010)Quality Engineering Project of Universities in Guangdong Province (KA220160216,KA220160149)。
文摘提出了一种基于改进YOLOv5n的肉羊攻击行为检测方法.在轻量级网络YOLOv5n的基础上,首先利用Ghost模块卷积替换传统卷积,减少网络参数和计算成本;其次在网络关键位置添加坐标注意力机制(Coordinate Attention,CA),用来增强网络通道信息和位置信息;然后引入EIOU(Efficient Intersection Over Union)损失函数,并进行改进,增强预测框准确性,减少误检情况;最后利用Ranger21优化器替代SGD(Stochastic Gradient Descent)优化器,缓解Ghost模块卷积提取特征不足,加快模型收敛速度.为了提升模型对不同天气泛化能力,对一半数据添加雪天、雨天和雾化3种天气情况.对比试验表明,改进模型是原YOLOv5n基础模型参数的83%,模型大小仅有3 Mb,而准确率提升0.7%,召回率提升1.3%,mAP(Mean Average Precision)提升0.5%,F1提升1.01%,同时优于其他主流基于YOLO轻量级网络模型.