基于YOLOv8-ABW的黄花成熟度检测方法  

YOLOv8-ABW based method for detecting Hemerocallis citrina Baroni maturity

作　　者：吴利刚 陈乐 刘泽鹏 武晔秋 马宇波 史建华 WU Ligang;CHEN Le;LIU Zepeng;WU Yeqiu;MA Yubo;SHI Jianhua(College of Mechanical and Electrical Engineering,Shanxi Datong University,Datong 037003,China;College of Coal Engineering,Shanxi Datong University,Datong 037003,China;College of Architecture and Surveying Engineering,Shanxi Datong University,Datong 037003,China;College of Physics and Electronic Science,Shanxi Datong University,Datong 037003,China)

机构地区：[1]山西大同大学机电工程学院,大同037003 [2]山西大同大学煤炭工程学院,大同037003 [3]山西大同大学建筑与测绘工程学院,大同037003 [4]山西大同大学物理与电子科学学院,大同037003

出　　处：《农业工程学报》2024年第13期262-272,共11页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然科学基金项目(12375050);山西省基础研究计划项目(202303021211330);大同市科技计划项目(2023015,2023006);山西大同大学基础科研基金项目(2022K1)。

摘　　要：为实现黄花成熟度的快速、高精度识别,针对其相似特征识别精确度低以及相互遮挡检测困难的问题,提出一种基于YOLOv8-ABW的黄花成熟度检测方法。该研究在特征提取网络中加入结合注意力机制的尺度特征交互机制(attention based intra-scale feature interaction, AIFI),更好地提取黄花特征信息,提高检测的精确度。在特征融合网络中,进一步采用加权的双向特征金字塔特征融合网络(bidirectional feature pyramid network, Bi FPN),实现更高层次的跨通道特征融合,有效减少通道中的特征冗余。此外使用WIoUv3作为损失函数,聚焦普通质量的锚框,提高模型的定位性能。试验结果表明:YOLOv8-ABW模型检测精确度为82.32%,召回率为83.71%,平均精度均值mAP@0.5和m AP@0.5:0.95分别为88.44%和74.84%,调和均值提升至0.86,实时检测速度为214.5帧/s。与YOLOv8相比,YOLOv8-ABW的精确度提高1.41个百分点,召回率提高0.75个百分点,mAP@0.5和m AP@0.5:0.95分别提升1.54个百分点和1.42个百分点。对比RT-DETR、YOLOv3、YOLOv5、YOLOv7模型,YOLOv8-ABW参数量最少,仅为3.65×10~6,且模型浮点运算量比YOLOv7少96.3 G。体现出YOLOv8-ABW模型能够在黄花成熟度检测中平衡检测精确度和检测速度,综合性能最佳,为黄花智能化实时采摘研究提供技术支持。Here,rapid and high-precision detection was proposed for Hemerocallis citrina Baroni maturity using YOLO v8-ABW.Current challenges were overcome on the similar features and mutual occlusions during recognition.The precision and efficiency of detection were improved to provide crucial technical support to intelligent real-time harvesting.Feature extraction and utilization were also greatly improved in the detection process of Hemerocallis citrina Baroni maturity.Attention-based Intra-scale Feature Interaction(AIFI)was incorporated into the feature extraction network.Feature information was interacted with and combined from various scales.The detection precision was enhanced to more effectively extract the information about Hemerocallis citrina Baroni.Specifically,the AIFI was used to weigh the features using the attention mechanism.More key areas were focused on extracting the features,in order to reduce the interference of noise and redundancy.Meanwhile,the scale feature interaction was used to fully utilize the feature information of different scales,thereby enhancing the precision and robustness of feature extraction.A weighted bidirectional feature pyramid feature fusion network(BiFPN)was used in the feature fusion network.This network structure was achieved in the complementarity and enhancement of various layers of feature information using cross-channel feature fusion.Compared with the traditional Feature Pyramid Network,BiFPN was used to retain more original feature information and fuse the features of different layers in a weighted manner,thus enriching the fused features.In addition,Bi FPN had effectively reduced the feature redundancy in the channels,thereby enhancing the speed and efficiency of detection.Moreover,WIoUv3 was used in the selection of the loss function.The loss function was specifically optimized for the standard quality anchor frames,in order to focus more on the location of targets during training.The WIoUv3 loss function was introduced to successfully enhance the localization performan

关 键 词：机器学习 模型 YOLOv8 黄花 深度学习 成熟度检测 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]