摘要
对杂草的检测是实现自动化杂草防治的重要先决条件,而基于深度学习的目标检测技术是实现杂草精准检测的有效手段,为克服田间复杂场景下杂草目标特征难以提取以致漏检误检突出的问题,论文提出了基于改进YOLOv5的杂草检测方法。该方法通过引入Swin Transformer模块增强对全局信息和上下文信息的提取能力,提升模型对杂草目标的感知度;结合卷积注意力机制,强化模型从大视野图像中关注小型目标的能力;对检测头进行解耦合,解决分类和定位兴趣区域不同引起的问题;选用EIOU损失函数实现更高精度的定位。在菜田伴生杂草数据集Ronin和开放植物表型数据集OPPD上的对比实验表明,改进模型的mAP达到92%,相比改进前的YOLOv5提升4.1%,较好地解决了精准杂草防治中的感知问题。
The detection of weeds is an important prerequisite for automatic weed control,and the object detection technology based on deep learning is an effective means to achieve accurate detection of weeds.In order to overcome the difficulty of extracting weed objects in complex field scenes,resulting in outstanding missed detection and false detection,this paper proposes a weed de-tection method based on improved YOLOv5.This method introduces the Swin Transformer module to enhance the extraction ability of global information and context information,and improve the model's perception of weed objects.Combined with the convolutional attention mechanism,the ability of the model to focus on small objects from large-field images is strengthened.Decoupling the de-tection head solves the problems caused by different regions of interest in classification and localization.The EIOU loss function is selected for higher accuracy positioning.Comparative experiments on the vegetable field associated weed dataset Ronin and the open plant phenotype dataset OPPD show that the mAP of the improved model reaches 92%,which is 4.1% higher than that of YOLOv5 before the improvement,which better solves the perception problem in precision weed control.
作者
秦立浩
董峦
张世豪
逄正钧
QIN Lihao;DONG Luan;ZHANG Shihao;PANG Zhengjun(College of Computer and Information Engineering,Xinjiang Agricultural University,Urumqi 830052)
出处
《计算机与数字工程》
2025年第8期2194-2199,2251,共7页
Computer & Digital Engineering
