基于YOLO v8n-seg-FCA-BiFPN的奶牛身体分割方法  

Segmentation Model of Cow Body Parts Based on YOLO v8n-seg-FCA-BiFPN

作　　者：张姝瑾 许兴时 邓洪兴 温毓晨 宋怀波[1,2] ZHANG Shujin;XU Xingshi;DENG Hongxing;WEN Yuchen;SONG Huaibo(College of Mechanical and Electronic Engineering,Northwest A&F University,Yangling,Shaanxi 712100,China;Key Laboratory of Agricultural Internet of Things,Ministry of Agriculture and Rural Affairs,Yangling,Shaanxi 712100,China)

机构地区：[1]西北农林科技大学机械与电子工程学院,陕西杨凌712100 [2]农业农村部农业物联网重点实验室,陕西杨凌712100

出　　处：《农业机械学报》2024年第3期282-289,391,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划项目(2023YFD1301800);国家自然科学基金项目(32272931)。

摘　　要：奶牛身体部位的精准分割广泛应用于奶牛体况评分、姿态检测、行为分析及体尺测量等领域。受奶牛表面污渍和遮挡等因素的影响,现有奶牛部位精准分割方法实用性较差。本研究在YOLO v8n-seg模型的基础上,加入多尺度融合模块与双向跨尺度加权特征金字塔结构,提出了YOLO v8n-seg-FCA-BiFPN奶牛身体部位分割模型。其中,多尺度融合模块使模型更好地提取小目标几何特征信息,双向跨尺度加权特征金字塔结构实现了更高层次的特征融合。首先在奶牛运动通道处采集奶牛侧面图像作为数据集,为保证数据集质量,采用结构相似性算法剔除相似图像,共得到1 452幅图像。然后对目标奶牛的前肢、后肢、乳房、尾部、腹部、头部、颈部和躯干8个部位进行标注并输入模型训练。测试结果表明,模型精确率为96.6%,召回率为94.6%,平均精度均值为97.1%,参数量为3.3×10^(6),检测速度为6.2 f/s。各部位精确率在90.3%~98.2%之间,平均精度均值为96.3%。与原始YOLO v8n-seg相比,YOLO v8n-seg-FCA-BiFPN的精确率提高3.2个百分点,召回率提高2.6个百分点,平均精度均值提高3.1个百分点,改进后的模型在参数量基本保持不变的情况下具有更强的鲁棒性。遮挡情况下该模型检测结果表明,精确率为93.8%,召回率为91.67%,平均精度均值为93.15%。结果表明,YOLO v8n-seg-FCA-BiFPN网络可以准确、快速地实现奶牛身体部位精准分割。The fine segmentation of cow body parts has significant applications in research fields such as cow body condition scoring,posture estimation,behavior recognition,and body measurement.Due to the limited practicality of existing segmentation methods for different cow body parts,an improved YOLO v8n-seg model named YOLO v8n-seg-FCA-BiFPN was proposed for cow body part segmentation tasks.The improved model added FCA channel attention mechanism to the YOLO v8n backbone feature extraction network to better extract the geometric feature information of small targets,and used repeated weighted bidirectional features in the network feature fusion layer.The BiFPN was used to achieve the purpose of increasing the coupling of features at each scale.In order to validate the model performance,side-view images of cows at the channel were collected for network training.To ensure the quality of the dataset,the structural similarity algorithm was used to remove similar redundant images,resulting in a total of 1452 images.LabelMe software was used to label the target cows,which were divided into eight parts,forelimbs,hindlimbs,udders,tails,belly,head,neck,and trunk,and was sent to the training model.The test results showed that the precision was 96.6%,the recall was 94.6%and the mean average precision was 97.1%,the parameters number was 3.3×106,and the detection speed was 6.2f/s.The precision of each part was from 90.3%to 98.2%,and the mean average precision was 96.3%.The YOLO v8n-seg-FCA-BiFPN network could realize accurate segmentation of various parts of dairy cows.Compared with the original YOLO v8n,the precision,recall and mean average precision of YOLO v8n-seg-FCA-BiFPN were 3.2 percentages points,2.6 percentages points and 3.1 percentages points higher than that of YOLO v8n-seg,respectively.The precision under occlusion was 93.8%,the recall value was 91.67%,and the mean average precision was 93.15%.The volume of the improved model remained unchanged and had strong robustness.Under occlusion,the precision was 93.8%,the recal

关 键 词：奶牛 身体部位分割 语义分割 FCABasicBlock BiFPN YOLO v8n 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]