Authors

Abstract

In a real-life breeding environment, fine-grained pig behavior detection is of great significance for health assessment, welfare monitoring, and intelligent management. However, issues such as dense target distribution, severe occlusion, subtle inter-class differences, class imbalance, and limited computing resources make it difficult to achieve both detection accuracy and computational efficiency. To address these problems, this paper proposes a lightweight pig behavior detection model WFE-YOLO based on YOLOv11, which is used to identify five typical behaviors: standing, lying down, eating, drinking, and chewing. This method conducts collaborative optimization at three levels: training sample distribution, feature representation, and detection head design. Specifically, a weighted sampling strategy is adopted to enhance the learning sufficiency of low-frequency behaviors; a lightweight gated feature extraction module is introduced to improve the fine-grained representation ability; an efficient detection head is designed to reduce structural redundancy and computational overhead. Experimental results show that on the data set of this paper, the Precision, Recall, and mAP@50 of WFE-YOLO reach 0.8154, 0.7803, and 0.8233 respectively; compared with YOLOv11n, the parameter size is reduced from 2.58M to 1.96M, GFLOPs is reduced from 6.3G to 4.4G, and FPS reaches 520.43. Under the experimental settings adopted in this paper, compared with several mainstream YOLO models, WFE-YOLO demonstrates a better balance between detection performance and model complexity, especially in low-frequency and easily confused behaviors such as Drink and Bite, and has a greater advantage. These results indicate that WFE-YOLO provides a lightweight, application-oriented solution for pig behavior monitoring in complex breeding environments, with strong potential for deployment on edge devices.

Abstract in English

在真实养殖环境中，细粒度猪行为检测对于健康评估、福利监测和智能化管理具有重要意义。然而，目标密集分布、严重遮挡、类间差异细微、类别不平衡以及计算资源受限，使得检测精度与计算效率难以兼顾。为解决这些问题，本文提出了一种基于 YOLOv11 的轻量化猪行为检测模型 WFE-YOLO，用于识别五类典型行为：站立、躺卧、进食、饮水和啃咬。该方法从训练样本分布、特征表示和检测头设计三个层面进行协同优化。具体而言，采用加权采样策略增强低频行为的学习充分性；引入轻量化门控特征提取模块以提升细粒度表示能力；设计高效检测头以减少结构冗余和计算开销。实验结果表明，在本文数据集上，WFE-YOLO 的 Precision、Recall 和 mAP@50 分别达到 0.8154、0.7803 和 0.8233；与 YOLOv11n 相比，参数量由 2.58M 降至 1.96M，GFLOPs 由 6.3G 降至 4.4G，FPS 达到 520.43。在本文采用的实验设置下，与多种主流 YOLO 模型相比，WFE-YOLO 在检测性能与模型复杂度之间表现出较优的平衡，尤其在 Drink 和 Bite 等低频且易混淆行为上更具优势。这些结果表明，WFE-YOLO 为复杂养殖环境下的猪行为监测提供了轻量化、面向应用的技术参考，并且具备终端部署的潜力。