基于改进YOLOv5的红花目标检测算法研究  

Research on Target Detection Algorithm of Safflower Balls Based on Improved YOLOv5

作　　者：陈金荣 许燕[1,3] 周建平 王小荣[2] Chen Jinrong;Xu Yan;Zhou Jianping;Wang Xiaorong(College of Mechanical Engineering,Xinjiang University,Urumqi 830047,China;Engineering Training Centre,Xinjiang University,Urumqi 830047,China;Agriculture and Animal Husbandry Robot and Intelligent Equipment Engineering Research Center of Xinjiang Uygur Autonomous Region,Urumqi 830047,China)

机构地区：[1]新疆大学机械工程学院,乌鲁木齐830047 [2]新疆大学工程训练中心,乌鲁木齐830047 [3]新疆维吾尔自治区农牧机器人及智能装备工程研究中心,乌鲁木齐830047

出　　处：《农机化研究》2025年第1期26-32,66,共8页Journal of Agricultural Mechanization Research

基　　金：新疆维吾尔自治区创新团队项目(2022D14002);新疆农机研发制造推广应用一体化项目(YTHSD2022-05)。

摘　　要：为实现农业非结构环境下采摘机器人对红花的准确识别,提出了一种基于改进YOLOv5的红花目标检测算法。将CBAM注意力机制嵌入到YOLOv5网络,提高了小尺寸目标物在高层次特征中的表现力;建立一种Alpha-IoU目标位置损失函数对原损失函数GIOU存在的梯度消失问题进行改进,提高了被遮挡红花的预测率,并通过在目标检测网络中增加分割检测模块,提高宽和高小于最低像素的小目标物检测精度,利用图像扩增数据集对改进后的YOLOv5算法进行训练,再分别与改进前后YOLOv5网络和Faster R-CNN网络在不同红花品种、不同自然光照情况、不同天气条件和不同遮挡情况下进行对比。试验结果表明:改进后的YOLOv5算法P值、R值分别为90.45%和0.90,对非结构环境下盛开期的未采摘红花mAP值达到94.48%,在不同影响因素下都可以准确识别出红花且置信度较高,可为红花采摘机器人自动化作业中的红花识别提供技术支持。In order to realize accurate recognition of safflower by picking robot in agricultural unstructured environment,a safflower target detection algorithm based on improved YOLOv5 was proposed.By embedding CBAM attention mechanism into YOLOv5 network,the expressiveness of small size objects in high-level features is improved.An Alpha-IoU target position loss function is established to improve the gradient disappearance problem existing in the original loss function GIOU,and the prediction rate of blocked red safflower is improved.By adding segmentation detection module into the target detection network.The detection accuracy of small objects with width and height less than the lowest pixel was improved,and the improved YOLOv5 algorithm was trained by using the image amplification data set.Then,the YOLOv5 network and Faster R-CNN network before and after the improvement were compared under different safflower varieties,different natural lighting conditions,different weather conditions and different occlusion conditions.The experimental results show that the P value and R value of the improved YOLOv5 algorithm are 90.45% and 0.90 respectively,and the mAP value of the unpicked safflower in the non-structural environment in blooming stage reaches 94.48%.Under different influencing factors,the algorithm can accurately identify safflower with high confidence,which can provide technical support for the automatic safflower picking robot recognition.

关 键 词：红花 目标检测 改进YOLOv5 数据增强 非结构环境 

分 类 号：S225[农业科学—农业机械化工程] TP391.4[农业科学—农业工程]