Authors

Abstract

Accurate individual identification is essential for precision donkey farming, but conventional methods based on manual records or physical tags are labor-intensive, inefficient, and vulnerable to tag loss or damage. To address these limitations, this study proposes MFW-YOLO11, a non-contact donkey face localization and individual identification model based on an improved YOLO11 framework. A total of 6,531 valid donkey face images were collected from a real farm environment and used to construct an individual identity dataset under diverse conditions, including different illumination, poses, occlusions, and backgrounds. In the proposed network, MANet is introduced into the backbone and head to strengthen fine-grained identity-related features, such as the eyes, nasal bridge, muzzle region, facial contour, and coat texture. A MANet-FasterCGLU composite module is further designed to adaptively filter effective facial responses and suppress interference from railings, donkey bodies, troughs, and complex backgrounds. In addition, a weighted feature union module is embedded in the neck to enhance the adaptive fusion of shallow texture details and deep semantic information. Experimental results show that MFW-YOLO11 achieves a precision of 90.7%, recall of 79.8%, mAP50 of 88.0%, mAP50–95 of 74.4%, FPS of 68.93, and GFLOPs of 6.3. Compared with the original YOLO11, the proposed model improves precision, recall, mAP50, and mAP50–95 by 6.5, 8.8, 6.2, and 6.4 percentage points, respectively, while maintaining real-time inference performance. These results indicate that MFW-YOLO11 provides an effective and practical solution for non-contact donkey individual identification in precision livestock management.

Abstract in Chinese

针对传统驴只个体识别方法依赖人工记录和物理标签、识别效率低且易受标签脱落影响等问题，提出一种基于改进YOLO11的非接触式驴脸定位与个体识别模型MFW-YOLO11。研究在真实驴场环境下采集并筛选6,531张有效驴脸图像，构建覆盖不同光照、姿态、遮挡和背景条件的个体身份数据集。模型在Backbone和Head中引入MANet增强眼部、鼻梁、口鼻区域、脸部轮廓和毛色纹理等细粒度身份特征表达；构建MANet-FasterCGLU复合模块以自适应筛选有效驴脸响应，并抑制栏杆、驴身、饲槽和复杂背景干扰；在Neck中嵌入WFU模块以增强浅层纹理细节和深层语义信息的加权融合。实验结果表明，MFW-YOLO11的Precision、Recall、mAP50、mAP50-95、FPS和GFLOPs分别达到90.7%、79.8%、88.0%、74.4%、68.93和6.3，较原始YOLO11分别提高6.5、8.8、6.2和6.4个百分点。结果表明，该模型可为精准畜牧管理中的驴只非接触式个体识别提供技术支撑。