基于牛脸和躯干综合信息的奶牛个体识别研究
Individual identification of dairy cows based on comprehensive face and trunk information
赵玲 1周桂红 1任力生1
作者信息
- 1. 河北农业大学 信息科学与技术学院 /河北省农业大数据重点实验室,河北 保定 071001
- 折叠
摘要
针对基于单一信息对奶牛个体身份识别精度低的问题,本文提出 1 种基于牛脸和躯干综合信息对奶牛个体身份识别的方法.在Mask R-CNN目标检测模型的基础上进行改进,将注意力机制模块引入到Mask R-CNN的ResNet50 特征提取网络的输出阶段,能够在图像通道和空间上增强奶牛身份信息.针对奶牛不同部位,本文对改进前后的Mask R-CNN模型分别基于牛脸、基于躯干以及基于牛脸和躯干综合信息进行了相关实验.实验结果表明,原始Mask R-CNN模型基于牛脸和躯干综合信息进行奶牛个体识别,比单独基于牛脸或躯干的识别精度提高 2.3%～3.7%.改进后的Mask R-CNN模型在自建奶牛图像数据集上的准确率达到了 93.63%,mAP值达到 92.16%,相较于原始Mask R-CNN,准确率提高了 2.92%,mAP值提高了 2.63%.本文方法能够实现对养殖场环境下奶牛个体身份的识别,可为奶牛的精准养殖提供技术支持.
Abstract
In order to increase the low accuracy in identifying individual dairy cows based on single information,a method based on the comprehensive information of the dairy cow's face and trunk was proposed.On the basis of the Mask R-CNN model,the Convolutional block attention module was introduced into the output stage of Mask R-CNN's ResNet50 feature extraction network,which could enhance the identity information of dairy cows in image channel and space.Based on different parts of dairy cows,this paper used the improved Mask R-CNN model to carry out relevant experiments with the information of dairy cow face,trunk and comprehensive information of dairy cow face and trunk.The experimental results showed that the comprehensive information of dairy cow face and trunk increased the recognition accuracy of the original Mask R-CNN model that was 2.3%to 3.7%higher than the face or trunk single information.The accuracy rate was 93.63%and mAP was 92.16%of improved Mask R-CNN model on the self-built data set of dairy cows,which has been increased by 2.92%and 2.63%respectively compared with the original Mask R-CNN model.The method in this paper could identify individual dairy cows in farms and provide technical support for precise breeding of dairy cows.
关键词
奶牛/牛脸和躯干/个体识别/Mask/R-CNN/注意力机制Key words
dairy cows/cow face and trunk/individual identification/Mask R-CNN/convolutional block attention引用本文复制引用
基金项目
河北省重点研发计划项目(19220119D)
出版年
2024
国家科技期刊平台
NSTL
万方数据