Authors

Abstract

Sheep face detection is critical for intelligent livestock management and breeding, yet existing models often struggle in complex farm scenarios due to inadequate multi-scale feature utilization and high computational demands. To address these challenges, this study proposes a lightweight multi-breed sheep face detection framework named YOLO-LSD (Lightweight Sheep Face Detection), achieving an optimal balance between detection accuracy and computational efficiency through multi-dimensional optimizations. At the feature enhancement level, the lightweight channel attention mechanism Efficient Channel Attention (ECA) is embedded into the backbone network to dynamically strengthen the channel responses of key facial features through local cross-channel interactions. Concurrently, Ghost convolution is introduced to replace traditional convolutional layers, leveraging feature redundancy mining technology to substantially reduce computational complexity while maintaining the ability to represent diverse facial features across sheep and goat breeds. To address the limited sample diversity in multi-breed datasets, a transfer learning strategy is employed, involving directional fine-tuning of breed-specific facial features based on large-scale pre-trained models to enhance the model's generalization ability across diverse sheep and goat varieties. Experimental results demonstrate that YOLO-LSD achieves a mAP@0.5 of 99.29% on a self-constructed multi-breed sheep face dataset, marking a 0.59% improvement over the baseline YOLOv11. Notably, the parameter count of YOLO-LSD is only 2.4×106, while achieving an inference speed of 60 FPS and 6.3 Flops. This study presents a high-precision, lightweight solution for intelligent livestock monitoring systems, offering practical insights for the deployment of multi-breed sheep face detection models in real-world farm applications.

Abstract in Chinese

绵羊面部检测是智能畜牧管理和养殖的关键，但由于多尺度特征利用不足和计算需求高，现有模型在复杂的农场场景中往往难以实现。为了解决这些挑战，本研究提出了一种轻量级的多品种羊人脸检测框架，名为YOLO-LSD (lightweight sheep face detection)，通过多维优化实现了检测精度和计算效率之间的最佳平衡。在特征增强层面，将轻量级通道注意机制ECA嵌入骨干网络，通过局部跨通道交互，动态增强关键面部特征的通道响应。同时，引入Ghost卷积来取代传统的卷积层，利用特征冗余挖掘技术大幅降低计算复杂性，同时保持表示绵羊和山羊品种不同面部特征的能力。为了解决多品种数据集样本多样性有限的问题，采用迁移学习策略，在大规模预训练模型的基础上对特定品种的面部特征进行定向微调，以提高模型在不同绵羊和山羊品种间的泛化能力。实验结果表明，YOLO-LSD在自构建的多品种绵羊面部数据集上的mAP@0.5达到了99.29%，比基线YOLOv11提高了0.59%。值得注意的是，YOLO-LSD的参数量仅为2.4×106，同时实现了60 FPS和6.3 Flops的推理速度。本研究提出了一种高精度、轻量级的智能牲畜监测系统解决方案，为在实际农场应用中部署多品种绵羊面部检测模型提供了实用的见解。