【摘要】 为高效准确地识别奶牛行为，提出一种改进YOLOv8的奶牛行为识别模型。使用EfficientNet框架替换骨干网络，增强模型特征提取能力；在颈部网络构建C2F—MLCA模块，增强提取信道和空间信息的能力，并融合局部和全局特征；使用共享卷积构建新的L—Head检测头，有效降低复杂度；使用Inner—IoU损失函数结合辅助边框动态调整，克服CIoU的局限性，提高模型的性能；采用信道剪枝来消除冗余参数，压缩模型结构，降低复杂度。改进YOLOv8s模型mAP相较于原模型提高2.2%，参数量减少74.7%，计算量减少59.8%，模型大小减少73.1%；相比于Faster R—CNN、YOLOv5s、YOLOv7和YOLOv8s模型，mAP分别提高8.7%、2.6%、0.1%和2.2%。应用模型统计分析奶牛日常行为比例与持续时间，突出监测奶牛行为对环境检测和评估奶牛状况的重要性，为奶牛自动化管理提供参考。更多还原

【Abstract】 To enhance the efficiency and accuracy of cow behavior recognition, we propose an improved YOLOv8—based model. The model replaces the backbone network with the EfficientNet framework to enhance feature extraction capabilities. A C2F—MLCA module is constructed in the neck network to enhance the ability to extract channel and spatial information and to integrate local and global features. A new L—Head detection head is built using shared convolution to effectively reduce complexity. The Inner—IoU loss function, combined with auxiliary bounding box dynamic adjustment, is employed to overcome the limitations of CIoU and improve the model′s performance. Channel pruning is adopted to eliminate redundant parameters, compress the model structure, and reduce complexity. Compared to the original YOLOv8s model, the improved model demonstrates a 2. 2% increase in mAP, a 74. 7% reduction in parameters, a 59. 8% decrease in computational load, and a 73. 1% reduction in model size. Compared to Faster R—CNN, YOLOv5s, YOLOv7, and YOLOv8s models, the mAP is improved by 8. 7%, 2. 6%, 0. 1%, and 2. 2%, respectively. The application of the model in analyzing the proportion and duration of daily cow behaviors highlights the importance of monitoring cow behavior for environmental detection and cow health evaluation, providing valuable insights for automated cow management.更多还原