【摘要】 哺乳期仔猪因体型小、生性好动、易被遮挡且聚集重叠，人工盘点效率低且易出错。现有方法在复杂猪场环境和频繁遮挡下难以平衡检测准确率与资源受限部署需求，增加了漏检和误检风险。为此，本文提出了一种基于YOLO v8n的轻量化仔猪检测模型GSD-YOLO。模型通过引入柔性非极大值抑制边界框交并操作和Inner-MPDIoU损失函数，优化边界框回归以降低误检和漏检率；并嵌入坐标注意力机制(Coordinate attention for efficient mobile network design, CA),增强了目标区域的特征表达能力，有效解决长程依赖问题。为实现嵌入式设备的高效部署，模型引入GhostNet模块优化特征提取和融合，减少通道间特征冗余的同时降低模型参数量。模型重构了一种轻量化的检测头Detect＿DG,在模型体积缩减18.48%的同时，进一步提升了检测精度。与YOLO v8n相比，GSD-YOLO的F1分数和平均精度分别提升1.0、0.6个百分点，参数量降低61.28%,帧率提高12.5%。GSD-YOLO在综合检测性能上优于YOLO v11等4种主流模型。结果表明，该模型在不同遮挡、重叠和光照下检测仔猪目标的准确率更优，且模型内存占有量较小，仅有2.6 MB。将GSD-YOLO部署到边缘计算设备Jetson Orin NX和安卓(Android)移动端，为实际应用中的仔猪检测提供了有效的技术支撑。更多还原

【Abstract】 In modern pig farming, the detection and counting of piglets presents a significant challenge due to their small size, high mobility, tendency to be occluded, and habit of clustering. Manual counting is not only time-consuming and labor-intensive but also prone to errors. To address these issues, a novel piglet detection algorithm, GSD-YOLO, was proposed based on YOLO v8n, which was designed to provide fast and accurate detection. The algorithm tackled the inherent difficulty of detecting piglets by introducing flexible non-maximum suppression(FNMS) for bounding box intersection handling and employing the Inner-MPDIoU loss function to optimize the bounding box regression mechanism. Additionally, a coordinate attention(CA) mechanism was integrated to enhance the feature representation of target areas, effectively resolving issues related to long-range dependencies. For efficient deployment on embedded devices, the model was further optimized for lightweight performance. Specifically, the backbone and neck components of the original model were replaced with the GhostNet module, which reduced both model parameters and feature redundancy in the channels. Furthermore, a lightweight detection head, Detect＿DG, was introduced, which reduced the overall model size by 18.48% while simultaneously improving detection accuracy. Compared with YOLO v8n, GSD-YOLO achieved improvements of 1.0 and 0.6 percentage points in F1-score and average precision(AP), respectively, while reducing the model size by 61.28% and improving the frame rate by 12.5%. In comprehensive detection performance, GSD-YOLO outperformed four widely used models, including YOLO v5. Experimental results demonstrated that the proposed model can rapidly and accurately detect piglets under challenging conditions, such as occlusion, overlap, and varying lighting environments. Moreover, with a compact memory footprint of only 2.6 MB. The deployment of GSD-YOLO on edge computing devices such as the Jetson Orin NX and Android mobile devices provided effective technological support for piglet counting and detection in practical applications.更多还原