[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"item-2780":3},{"id":4,"title":5,"url":6,"summary":7,"summary_zh":8,"content":8,"source_name":9,"source_url":6,"published_at":10,"category":11,"cover_url":8,"hotness":12,"is_selected":13,"score":14,"score_detail":15,"sources":23,"tags":27,"view_count":33,"doi":34,"paper":35,"created_at":58},2780,"基于深度学习数据增强技术的肉牛行为识别 与养殖状态评估研究","https:\u002F\u002Fdoi.org\u002F10.65436\u002Fhssj.v1i9.50","摘 要：肉牛行为的准确识别是智慧养殖中个体状态监测和精细化管理的重要基础。针对实际养殖场景中光照变化、行为姿态差异以及训练样本规模对识别性能影响不明确等问题，本文以9类肉牛典型行为图像为研究对象，构建基于深度学习骨干网络与组合数据增强的行为识别流程。数据集包含饮水、进食、争斗、舔舐、躺卧、爬跨、探究、站立和行走9类行为，覆盖低光、正常光照、高光和黑暗4种光照条件，共4500张图像。首先，在统一训练条件下比较ResNet-50、DenseNet-121、EfficientNet-B0和Swin-T四种骨干网络。实验结果表明，EfficientNet-B0的测试准确率为98.37%±1.36%，参数量为4.02 M，MACs为0.41 G；其准确率较ResNet-50高0.07个百分点，较DenseNet-121和Swin-T均高0.59个百分点，同时具有四种网络中最低的参数量与计算量，因此选其作为后续实验的固定骨干。进一步设置不同训练样本规模实验与数据增强消融实验，以分析样本量和翻转、颜色抖动、旋转、随机擦除等增强操作的影响。实验二结果表明，当每类训练样本由每类10张增至全量训练集（每类400张）时，测试准确率由69.70%±0.76%提升至98.37%±1.36%，训练样本规模对识别性能影响显著，且随样本量增大边际收益总体呈递减趋势。数据增强消融实验表明，逐项去除翻转、颜色抖动、旋转和随机擦除后，测试准确率较完整增强基线（98.37%±1.36%）分别下降2.74、3.11、2.89和2.96个百分点，验证了各增强操作对复杂光照与姿态条件下识别性能的贡献，其中颜色抖动的作用最为显著。研究可为复杂光照条件下肉牛行为的自动识别与智能养殖监测提供参考。 基金项目：安徽省教育厅优秀青年教师培育项目（YQYB2026117）；安徽省智慧农业技术与装备重点实验室开放基金（AEC2026016；AEC2025009）；安徽省教育厅自然科学研究项目（重大）（2024AH040217）；省级质量工程-教学研究项目（2023jyxm1009；2023xjzlts117；2023sdxx145）;基于人工智能下的家禽康养与疾病预警技术研究（S202413620055）；基于自监督学习与对抗增强的小样本信号的智能感知研究（2025SK019）；基于深度学习的生猪识别与康养研究（S202513620083）",null,"人文与社会研究学报","2026-09-17T00:00:00Z","论文",25,false,78,{"impact":16,"substance":17,"depth":18,"authority":19,"freshness":20,"relevant":21,"comment":22},16,22,18,12,10,1,"论文以4500张多光照肉牛行为图像系统比较四种骨干网络并完成样本规模与数据增强消融实验，方法扎实、结论可靠，对复杂光照下智能养殖监测有直接参考价值，值得进入每日精选。",[24,25],{"name":9,"url":6},{"name":9,"url":26},"https:\u002F\u002Fdoi.org\u002F10.65436\u002Fhssj.v1i9.48",[28,29,30,31,32],"农业人工智能","深度学习","智慧养殖","数据增强","肉牛行为识别",0,"10.65436\u002Fhssj.v1i9.50",{"doi":34,"openalex_id":36,"authors":37,"venue":9,"cited_by_count":33,"oa_url":6,"card":51,"direction":55,"ingested_from":57},"W7213445826",[38,40,42,44,46,48],{"name":39,"orcid":8},"孙帅",{"name":41,"orcid":8},"赵永才",{"name":43,"orcid":8},"李慢",{"name":45,"orcid":8},"郜静茹",{"name":47,"orcid":8},"关曼玉",{"name":49,"orcid":50},"Yanbo Li","https:\u002F\u002Forcid.org\u002F0000-0001-5144-2309",{"tldr":52,"method":53,"finding":54,"direction":55,"opportunity":56},"比较四种骨干网络并做数据增强消融，实现9类肉牛行为识别与养殖状态评估。","4500张9类行为图像，ResNet-50、DenseNet-121、Effic","EfficientNet-B0最优（98.37%，4.02M参数）；样本量影响显著，颜色抖动贡献最大","智慧农业 \u002F 农业物联网","可探索轻量模型在边缘设备实时部署及跨牧场域适应，并引入时序视频行为识别。","openalex","2026-09-17T23:30:22.897005Z"]