摘要
为进一步提高奶牛体况自动评分精度,构建了一种基于点云凸包距离的三维结构特征图,将其作为EfficientNet深度学习网络的输入,可实现奶牛体况自动评分误差在0.25以内识别的准确率提升。首先,对获取的奶牛背部深度图像进行预处理,得到含有主要体况信息从奶牛腰角骨到臀骨区域的点云;其次,对点云进行体素化和凸包化,计算每个外围体素到最近凸包面之间的距离,并投影至X-Y平面上,得到结构特征图;构建EfficientNet网络分类模型,采用鲸鱼优化算法(Whale optimization algorithm,WOA)对其缩放系数进行优化;最后,利用77头奶牛的5 119幅深度图像对模型进行训练、验证与测试,数据集比例为5∶3∶2。结果表明,奶牛体况评分(BCS)范围在2.25~4.00内,测试集中EfficientNet模型精准识别的图像达到73.12%,BCS识别误差在0.25和0.50以内的图像占比分别为98.6%和99.31%,平均识别速率为3.441 s/f,识别效果优于MobileNet-V2、XceptionNet和LeNet-5等模型。该方法可实现规模化养殖场中奶牛个体体况的无接触评定,具有精度高、适用性强、成本低等特点。
Depth images are increasingly used to detect the body condition of dairy cows to make breeding management decisions.The scoring method of individual dairy cows body condition based on deep learning can further improve the degree of automation of dairy cow body condition image analysis.In order to realize the accurate recognition of the individual body condition of dairy cows without contact,high accuracy and strong applicability based on depth images in the actual breeding environment,a body condition scoring method was proposed based on deep learning and point cloud convex hulling features.Firstly,the acquired back depth image of the cow was preprocessed,included target extraction,target rotation,and acquired hindquarters images to obtain a back point cloud of the cow,containing the main body condition information.And then the hindquarters point cloud was voxelized and the convex hull feature image was obtained.In order to represent the fat and thin degree of different cows,and finally build a variety of convolutional neural network classification models,accuracy rate and average F1 value was used to optimize the model to further improve the accuracy of individual body condition recognition of dairy cows.The test results showed that the EfficientNet network can effectively identify the body condition of cows with a BCS value in the range of 2.25~4.00.The image account of recognition accuracy errors of 0.25 and 0.50 were 98.6%and 99.31%,respectively.The average F1 value was 98%and 99%,and the average recognition rate was 3.441 s/f.Compared with the MobileNet V2,XceptionNet,and LeNet 5 network models,the above indicators of the proposed method were better.The method can realize the non-contact assessment of the individual body condition of dairy cows in the breeding farm,and had the characteristics of high accuracy,strong applicability,and low cost.
作者
赵凯旋
刘晓航
姬江涛
ZHAO Kaixuan;LIU Xiaohang;JI Jiangtao(College of Agricultural Equipment Engineering,Henan University of Science and Technology,Luoyang 471003,China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2021年第5期192-201,73,共11页
Transactions of the Chinese Society for Agricultural Machinery
基金
国家重点研发计划项目(2019YFE0125600)
国家自然科学基金项目(32002227)
河南省科技攻关项目(192102110089)。
