基于改进YOLOv8-Seg的苹果单枝条花序疏除方法  

Thinning apple inflorescence at single branch level using improved YOLOv8-Seg

作　　者：司永胜[1] 孔德浩 王克俭[1] 刘丽星 杨欣[2] SI Yongsheng;KONG Dehao;WANG Kejian;LIU Lixing;YANG Xin(College of Information Science and Technology,Hebei Agricultural University,Baoding 071001,China;College of Mechanical and Electrical Engineering,Hebei Agricultural University,Baoding 071001,China)

机构地区：[1]河北农业大学信息科学与技术学院,保定071001 [2]河北农业大学机电工程学院,保定071001

出　　处：《农业工程学报》2024年第14期100-108,共9页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家现代农业产业技术体系项目(CARS-27)。

摘　　要：针对苹果疏花作业中无法自动识别枝干以及缺乏花序疏除方法,该研究提出了一种适用于现代果园的苹果树单枝条花序疏除方法。首先,对YOLOv8-Seg模型进行改进:在Backbone部分添加GCT(Gaussian context transformer)模块,通过引入全局上下文信息和调整通道的重要性,提高模型对遮挡目标的分割性能;在对应3个检测头的Neck部分的C2f模块内部增加EMA(efficient multi-scale attention)机制,通过并行子网结构和跨空间信息聚合更好地关注多尺度特征。其次,使用改进YOLOv8-Seg模型对单枝条中的花苞、花序、开放花朵和花枝四类目标进行实例分割。最后,基于分割结果应用多项式拟合曲线表征花枝,并计算花序间距离实现花序疏除。结果表明,改进的YOLOv8s-Seg模型在自建数据集mask水平的精确率、召回率和mAP分别达到了89.9%、89.5%和91%,比原模型分别提升了6.5、4.1和5.8个百分点。与主流分割模型Mask R-CNN,YOLACT,SOLOv2进行对比,mask水平的mAP分别高出10.8、12.3和9.1个百分点。花序疏除决策结果与人工决策结果对比误差不超过10%。该方法可应用于单枝条水平上的花序疏除任务,为苹果智能疏花提供技术支持。This study aims to realize the intelligent thinning operation on the apple tree.An inflorescence thinning was also proposed for apple trees at the single-branch level in modern orchards.Firstly,the YOLOv8-Seg model was improved as follows.The GCT(Gaussian Context Transformer,GCT)module was added to Backbone.The segmentation performance of the occluded targets was improved to introduce the global context information for the importance of channels.Efficient Multi-Scale Attention(EMA)mechanism was added inside the C2f module.The Neck portion of the three detection heads was better focused on the multi-scale features using parallel subnetwork structure and cross-space information aggregation.Secondly,the improved YOLOv8-Seg model was used to segment instances for four types of targets:buds,inflorescences,open flowers,and flowering branches at the single-branch level.Finally,the polynomial fitting curves were used to represent the flower branches on images after instance segmentation.At the same time,the distance between inflorescences was calculated for the inflorescence thinning.600 images of three apple varieties were collected,including Fuji,Gala,and Ruixue.An image dataset of the flower branch was then created using these images.The segmentation model and thinning decision were tested using the above dataset.The training was then conducted on the improved YOLOv8-Seg model.All losses were tended to stabilize.Then the model was converged after 1000 epochs.The mAP gradually increased to eventually stabilize during training.There was a small fluctuation of mAP during overall training.The model was then converged to the local optimal solution during training,indicating a relatively stable performance.The ablation experiment was used to verify the two improvements of the model,namely the GCT module and the EMA mechanism.The results showed that both improvements significantly enhanced the original model.50 images of Fuji,Gala,and Ruixue apple varieties were randomly selected to verify the robustness of the improved YOLO

关 键 词：图像处理 模型 苹果疏花 深度学习 实例分割 YOLOv8-Seg 花序疏除 

分 类 号：S24[农业科学—农业电气化与自动化] TP391.41[农业科学—农业工程]