EUVIMEDEUVIMED entdecken
Crossref

LCCD-Net: high-performance compositional computer vision framework for sheep posture recognition

Peiyun Xue, Jiafei Li, Jiayi Yang, Jian’an Zhang, Mingyu Cai, Shengnan Li, Ning Liu, Jiangong Li

Measurement Science and Technology · 26.06.2026 · 37 (26) · 266002

DOI: 10.1088/1361-6501/ae7aab ↗

Kurz gefasst

Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 25.06.2026. Verfasst von Peiyun Xue und 7 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben. Gefördert von: National Key Research and Development Program of China, National Key Laboratory of Animal Nutrition and Feeding.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

To address the issue of decreased behavior recognition accuracy in intensive indoor sheep farming scenarios caused by occlusion and confusingly similar postures (such as kneeling and sitting), this paper proposes a precision recognition framework based on a localization-classification collaborative decoupling strategy (LCCD-Net). The framework consists of two core phases: (1) In the localization phase, using YOLOv8n as the baseline, by embedding a global context block (GCBlock) into the C2f unit, adopting the Wise-IoU loss function, and introducing the dynamic upsampling operator (DySample), the enhanced localization architecture YOLOv8-GCWD was constructed, which significantly improves the perception capability for densely occluded targets; (2) In the classification phase, based on the ResNet framework, a side-branch multi-scale fusion mechanism is proposed to aggregate shallow texture and contour features, and a Last Block Fine-Tuning strategy is designed to capture subtle limb geometric changes, establishing the independent high-resolution classification network ResNet-LS, which effectively solves the posture misjudgment problem caused by multi-task feature coupling in single-stage models. Experimental results show that LCCD-Net achieves 93.1% mAP@0.5 in the object localization task, an increase of 4.0% over the YOLOv8n baseline, effectively improving the recall rate of occluded targets; meanwhile, it reaches 98.6% accuracy in the posture classification task, significantly outperforming the single-stage YOLOv8n and traditional two-stage methods. With a total parameter count of 31.7 M and an inference speed of 45 FPS, the framework enables real-time deployment on edge computing platforms.

Förderung

  • National Key Research and Development Program of China
  • National Key Laboratory of Animal Nutrition and Feeding

Zitieren

APA

Xue, P., Li, J., Yang, J., Zhang, J., Cai, M., Li, S., Liu, N., & Li, J. (2026). LCCD-Net: high-performance compositional computer vision framework for sheep posture recognition. Measurement Science and Technology, 37(26), 266002. https://doi.org/10.1088/1361-6501/ae7aab

Vancouver

Xue P, Li J, Yang J, Zhang J, Cai M, Li S, et al. LCCD-Net: high-performance compositional computer vision framework for sheep posture recognition. Measurement Science and Technology. 2026;37(26):266002. doi:10.1088/1361-6501/ae7aab

BibTeX
@article{Xue2026,
  author = {Xue, Peiyun and Li, Jiafei and Yang, Jiayi and Zhang, Jian’an and Cai, Mingyu and Li, Shengnan and Liu, Ning and Li, Jiangong},
  title = {LCCD-Net: high-performance compositional computer vision framework for sheep posture recognition},
  journal = {Measurement Science and Technology},
  year = {2026},
  volume = {37},
  number = {26},
  pages = {266002},
  doi = {10.1088/1361-6501/ae7aab}
}
Kostenlos starten →