Sheep-YOLO: improved and lightweight YOLOv8n for precise and intelligent recognition of fattening lambs' behaviors and vitality statuses

文献类型： 外文期刊

第一作者： Zhang, Mengjie

作者： Zhang, Mengjie;Hong, Dan;Wu, Jiabao;Zhu, Yanfei;Zhang, Xiaoshuan;Luo, Hailing;Zhao, Qinan

作者机构：

关键词： Fattening lambs; Computer vision; Deep learning; Improved and lightweight YOLOv8n

期刊名称：COMPUTERS AND ELECTRONICS IN AGRICULTURE （ 影响因子：8.9； 五年影响因子：9.3 ）

ISSN： 0168-1699

年卷期： 2025 年 236 卷

页码：

收录情况： SCI

摘要： Fattening lambs' behaviors and vitality statuses reflect health status and animal welfare directly. However, current detection methods for these aspects primarily rely on manual observation, which is inefficient. This paper aims to achieve precise and intelligent recognition of fattening lambs' behaviors and vitality statuses. Through the comprehensive literature review, field observations, and consultations with experts, four pivotal behaviors (running, sleeping, socialization, and wandering), alongside four vitality statuses (excitement, hungry, lethargy, and normal) are identified as crucial indicators that mirror the health status of fattening lambs. A visual system is established to collect and construct a dataset of fattening lambs, and a five-fold cross-validation method is employed to determine the basic model. Furthermore, an improved lightweight YOLOv8n model, named SheepYOLO, is developed in this paper. Sheep-YOLO utilizes the FasterNet network to reduce model complexity and incorporates the Mixed Local Channel Attention (MLCA) module, SIoU loss function, and Content-Aware ReAssembly of FEatures (CARAFE) to enhance the model's adaptability to varying breeding densities and lighting conditions. To validate the effectiveness of the optimization strategies, comparison experiments are conducted in this paper. The experimental results show that Sheep-YOLO achieved Params of 1.905 M and GFLOPs of 5.5 G, with a high mAP0.5 of 96.1 % on the test set and a detection speed of 79.317 FPS(12.61 ms/per image). Compared to the basic YOLOv8n, Sheep-YOLO achieves a 36.6 % reduction in Params and a 30 % decrease in GFLOPs, while maintaining higher mAP0.5 and faster detection speed. Besides, Sheep-YOLO surpasses widely used algorithms such as YOLOv3-tiny, YOLOv5n, YOLOv10n, RT-DETR, and Faster-RCNN regarding both lightweight performance and precision. Therefore, this study provides potential technical support for precise and intelligent recognition of the behaviors and vitality statuses of fattening lambs, contributing to the health monitoring and early disease prediction of sheep.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序