摘要
奶牛体况评分是评价奶牛产能与体态健康的重要指标。目前,随着现代化牧场的发展,智能检测技术已被应用于奶牛精准养殖中。针对目前检测算法的参数量多、计算量大等问题,以YOLO v5s为基础,提出了一种改进的轻量级奶牛体况评分模型(YOLO-MCE)。首先,通过2D摄像机在奶牛挤奶通道处采集奶牛尾部图像并构建奶牛BCS数据集。其次,在MobileNetV3网络中融入坐标注意力机制(Coordinate attention,CA)构建M3CA网络。将YOLO v5s的主干网络替换为M3CA网络,在降低模型复杂度的同时,使得网络特征提取时更关注于牛尾区域的位置和空间信息,从而提高了运动模糊场景下的检测精度。YOLO v5s预测层采用EIoU Loss损失函数,优化了目标边界框回归收敛速度,生成定位精准的预测边界框,进而提高了模型检测精度。试验结果表明,改进的YOLO v5s模型的检测精度为93.4%,召回率为85.5%,mAP@0.5为91.4%,计算量为2.0×109,模型内存占用量仅为2.28 MB。相较原始YOLO v5s模型,其计算量降低87.3%,模型内存占用量减少83.4%,在保证模型复杂度较低与实时性较高的情况下,实现了奶牛体况的高效评分。此外,改进的YOLO v5s模型的整体性能优于Faster R-CNN、SDD和YOLO v3目标检测模型。本研究为奶牛体况评分商业化提供理论基础和研究思路,为奶牛养殖业提供了智能化解决方案。
Cow body condition score is an important indicator to evaluate the productivity and physical health of cows.At present,with the development of modern farming,intelligent detection technology has been applied to precision farming of dairy cows.In view of the problems of large number of parameters and large calculation of the current detection algorithm,an improved lightweight cow body condition scoring model(YOLO MCE)was proposed based on YOLO v5s.Firstly,a 2D camera was used to acquire the cow tail images at the cow milking lane,and those images were filtered to obtain the final BCS dataset.Secondly,the coordinate attention(CA)mechanism was integrated into the MobileNetV3 network to build the M3CA network,which was used to replace the YOLO v5s backbone network to reduce the complexity of the model,and make the network feature extraction pay more attention to the location and spatial information of the cow tail area.Finally,the EIoU Loss function was used in the prediction layer of YOLO v5s to optimize the regression convergence speed of the target bounding box and generate a prediction bounding box with accurate positioning.The experimental results showed that the improved YOLO v5s model had a detection precision of 93.4%,a recall rate of 85.5%,an mAP@0.5 of 91.4%,a FLOPs of 2.0×109,and a model size of 2.28 MB.Compared with the original YOLO v5s model,the FLOPs and model size of YOLO MCE were reduced by 87.3%and 83.4%,respectively,which further showed that the proposed method can achieve efficient scoring of cow body conditions under the condition of low model complexity and high real⁃time performance.In addition,the overall performance of the improved YOLO v5s model was superior to that of the Fast R CNN,SDD and YOLO v3 object detection models.The research result can provide a theoretical basis and research ideas for the commercialization of dairy cow body condition scoring,and offer a research direction for the application of intelligent algorithms.
作者
黄小平
冯涛
郭阳阳
梁栋
HUANG Xiaoping;FENG Tao;GUO Yangyang;LIANG Dong(School of Internet,Anhui University,Hefei 230039,China;National Engineering Research Center for Agro-Ecological Big Data Analysis and Application,Hefei 230039,China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2023年第6期287-296,共10页
Transactions of the Chinese Society for Agricultural Machinery
基金
安徽省自然科学基金项目(1908085QF284)
安徽省教育厅自然科学基金项目(KJ2021A0024)。
关键词
奶牛体况评分
轻量化
目标检测
注意力机制
body condition score
lightweight
target detection
attentional mechanism
