摘要
猪只计数是大规模养猪场在养殖过程中的一项非常重要的工作.在复杂的猪圈环境中,由于猪的应激反应和诸多的障碍物,想要进行准确且自动化的计数是一个非常困难的工作.该文提出了一种基于YOLOv8的猪只数量目标检测模型(SCS-YOLOv8),旨在提高养殖场中猪只计数的准确性和效率.首先,该模型在第5层的C2f层融入Swin Transformer模块,增强模型的特征提取能力.同时,在第10层增加卷积块注意力模块(CBAM)注意力机制,增强模型对小目标和遮挡目标的检测能力,提升模型的鲁棒性.此外,还引入了简化的空间金字塔池化快速(SimSPPF)模块降低计算量,提高了推理速度,采用基于动态非单调聚焦机制的边界框回归(WIoU)损失函数更好地处理猪只的遮挡问题和目标较小的问题,提高模型在复杂场景下的检测性能,并结合软非极大值抑制方法(Soft-NMS)防止两个重合度过高的目标漏检.实验证明,该模型在自制数据集和科大讯飞公开数据集上均取得了优异的性能,其中在自有数据集上的mAP50-95值达到了77.2%,较初始YOLOv8x提高了4.8%,相较于其他YOLO模型都有不同程度的提升.同时该模型在科大讯飞的数据集上也有着不错的表现,证明了其良好的泛用性和鲁棒性.
Accurate and automated pig counting is a critical task in large-scale pig farms,yet it remains challenging due to the complex pen environments,pigs′stress responses,and numerous obstacles.This paper proposes a novel pig counting target detection model,SCS-YOLOv8,based on the YOLOv8 framework,to enhance counting accuracy and efficiency.The model integrates a Swin Transformer module into the C2f layer at the 5th layer to strengthen feature extraction capabilities.Additionally,a Convolutional Block Attention Module(CBAM)is incorporated at the 10th layer to improve detection of small and occluded targets,thereby enhancing model robustness.A Simplified Spatial Pyramid Pooling Fast(SimSPPF)module is introduced to reduce computational load and increase inference speed.The model employs a boundary box regression loss function based on the Dynamic Non-monotonic Focusing Mechanism(WIoU)to address occlusion and small target issues effectively.Moreover,Soft Non-Maximum Suppression(Soft-NMS)is utilized to prevent omission of highly overlapping targets.Experiments demonstrate that SCS-YOLOv8 achieves an mAP50-95 value of 77.2%on a self-made dataset,representing a 4.8%improvement over the original YOLOv8x and outperforming other YOLO models.The model also exhibits strong generalizability and robustness on the publicly available iFLYTEK dataset.
作者
欧阳建权
唐欢容
鲁嘉雄
OUYANG Jianquan;TANG Huanrong;LU Jiaxiong(School of Computer Science,Xiangtan University,Xiangtan 411105,China)
出处
《湘潭大学学报(自然科学版)》
2025年第5期20-31,共12页
Journal of Xiangtan University(Natural Science Edition)
基金
湖南省重点研发项目(NO.2023GK2057)
中央引导地方科技发展资金项目(2024ZYQ135)
湘潭大学研究生创新创业项目(XDCX2024Y283)。
