| 摘 要: 针对现有非接触式牛个体图像识别模型体积大、参数多、资源占用较大等问题,提出了一种基于改进YOLOv5模型的轻量级牛个体图像识别模型(Light YOLO Net,LY-Net)。将YOLOv5模型的主干网络替换为轻量级网络Ghost Net,并采用CARAFE(轻量级通用上采样算子),减少网络参数,实现网络轻量化;采用Focal-EIoULoss作为损失函数,加速收敛并提高了速度。采用甘肃省张掖市某养殖场的30头牛,共6 775幅牛个体图像作为样本数据,进行模型的训练、验证、测试。实验结果表明:LY-Net模型对牛个体的识别精确率约为99.6%,召回率约为99.5%。该模型能够在对牛个体图像高效且准确识别的同时,实现模型的小型化、轻量化。 |
| 关键词: :YOLOv5;牛个体;图像识别;轻量级 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 甘肃农业大学盛彤笙创新基金项目(GSAU-STS-2021-16);甘肃农业大学青年导师基金项目(GAU-QDFC-2021-18);甘肃省自然科学基金项目(20JR5RA023) |
|
| An Improved YOLOv5-based Method for Image Recognition of Cattle Individual |
|
LIU Qiwei1, GUO Xiaoyan1, LI Chunbin2, YANG Daohan2
|
(1.School of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China;
2.College of Resources and Environment, Gansu Agricultural University, Lanzhou 730070, China)
lqw18394511172@163.com; guoxy@gsau.edu.cn; licb@gsau.edu.cn; 529785408@qq.com
|
| Abstract: Aiming at the current research problems of the existing non-contact cattle individual image recognition models, such as large volume, multiple parameters and large resource occupation, this paper proposes a lightweight cattle individual image recognition model ( Light YOLO Net, LY-Net) based on improved YOLOv5 model. The backbone network of the YOLOv5 model is replaced by the lightweight network Ghost Net, and CARAFE ( a lightweight universal up-sampling operator) is used to reduce network parameters and realize network lightweight. Focal-EIoU Loss is used as the loss function to accelerate the convergence and improve the speed. A total of 6 775 individual cattle images of 30 cattle individuals on a farm in Zhangye City, Gansu Province are used as sample data for training, validation, and testing of the model. The experimental results show that the precision rate of LY-Net model for cattle individual recognition is about 99.6% , and the recall rate is about 99.5% . The proposed model can realize the miniaturization and lightweight of the model while effectively and accurately recognizing individual cattle images. |
| Keywords: YOLOv5; cattle individual; image recognition; lightweight |
