摘要
针对羊只个体差异较小,相似度高难以辨别,远距离识别准确率不高等问题,本文基于MobileFaceNet网络提出了一种融合空间信息的高效通道注意力机制的羊脸识别模型,对羊只进行非接触式识别。该研究基于YOLO v4目标检测方法生成羊脸检测器,以构建羊脸识别数据库;在MobileFaceNet的深度卷积层和残差层中引入融合空间信息的高效通道注意力(ECCSA),以增加主干特征的提取范围,提高识别率,并采用余弦退火进行动态学习率调优,最终构建ECCSA-MFC模型,实现羊只个体识别。试验结果表明,在羊脸检测上,基于YOLO v4的羊脸检测模型准确率可达97.91%,可以作为脸部检测器;在羊脸识别上,ECCSA-MFC模型在开集验证中识别率可达88.06%,在闭集验证中识别率可达96.73%。该研究提出的ECCSA-MFC模型在拥有较高识别率的同时更加轻量化,模型所占内存仅为4.8 MB,可为羊场智慧化养殖提供解决方案。
The difference between sheep is small,the similarity is high,it is difficult to distinguish,and the accuracy of long-distance recognition is not high.To solve that,a sheep face recognition model with efficient channel attention mechanism integrating spatial information was proposed to recognize sheep non-contact.The model was based on MobileFaceNet network.The research generated sheep face detector based on YOLO v4 target detection method was used to construct sheep face recognition database.An efficient channel attention integrating spatial information was introduced into the deep convolution layer and residual layer of MobileFaceNet to increase the extraction range of trunk features and improve the recognition rate.Cosine annealing was used to optimize the dynamic learning rate,and finally ECCSA-MFC model was built to realize sheep individual recognition.The experimental results showed that the accuracy of the sheep face detection model based on YOLO v4 can reach 97.91%and can be used as a face detector.In sheep face recognition,the recognition rate of ECCSA-MFC algorithm can reach 88.06%in open set verification and 96.73%in closedset verification.The proposed ECCSA-MFC model had higher recognition rate and lighter weight.The model size was only 4.8MB,which can provide a solution for intelligent breeding in sheep farm.
作者
张宏鸣
周利香
李永恒
郝靳晔
孙扬
李书琴
ZHANG Hongming;ZHOU Lixiang;LI Yongheng;HAO Jinye;SUN Yang;LI Shuqin(College of Information Engineering, Northwest A&F University, Yangling, Shaanxi 712100, China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2022年第5期267-274,共8页
Transactions of the Chinese Society for Agricultural Machinery
基金
国家重点研发计划项目(2020YFD1100601)。