基于改进YOLO v8n模型的散养蛋鸡个体行为识别方法与差异分析

Individual Behavioral Identification and Differential Analysis of Free-range Laying Hens Based on Improved YOLO v8n Model

摘要:

家禽行为与其生理状态密切相关，可利用行为数据对家禽健康状况进行评估。统计个体行为数据需要进行蛋鸡行为识别和个体身份识别，针对行为识别过程中，蛋鸡体型小、聚集遮挡，养殖环境光照变化等因素导致的蛋鸡有效特征表达不足，个体行为识别效果不理想问题，基于YOLO v8n网络构建行为识别模型，同时融合ODConv、GhostBottleneck、GAM注意力和Inner-IoU结构，通过减少图像特征丢失，放大全局交互信息，融合跨阶段特征，增强特征提取及泛化能力对模型进行改进，提升了蛋鸡采食、饮水、站立、整理羽毛、俯身搜索5种行为的识别精度。同时基于YOLO v8n模型构建了个体身份识别网络，并通过引入MobileNetV3模块对个体身份识别网络模型进行优化，提升了个体行为数据统计效率。试验结果表明，优化后行为识别模型对采食、饮水、站立、整理羽毛、俯身搜索行为识别平均精度(AP)分别达到94.4%、93.0%、90.7%、91.7%、86.9%，平均精度均值(mAP)达到91.4%，与YOLO v5n、YOLO v6n、YOLO v7-tiny、YOLO v8n相比，平均精度均值(mAP)分别提高4.8、4.1、5.5、3.5个百分点；个体身份识别模型参数量和运算量与YOLO v8n模型相比，减少1.9651×106和6.1×109。通过分析蛋鸡行为数据发现，行为数据与温度及蛋鸡个体本身有关，温度降低时，采食、站立次数增加，饮水次数减少，整理羽毛、俯身搜索次数几乎无变化，相同温度下，不同蛋鸡个体的行为数据差异较大，且差异值与蛋鸡体型有关。试验结果为依据行为数据评判蛋鸡健康状况、养殖场精准养殖及蛋鸡个体优选奠定了基础。

Abstract:

Poultry behavior is closely related to its physiological state, and behavioral data can be used to assess the health status of poultry. Statistical individual behavioral data is needed for laying hen behavioral identification and individual identification, to address the behavioral identification process, laying hen body size was small, aggregation of shade, breeding environment lighting changes and other factors resulting in the laying hen effective features expression was insufficient, individual behavioral identification effect was not ideal problem, based on the YOLO v8n network to build behavioral identification model, while fusing ODConv, GhostBottleneck, GAM attention and Inner-IoU structure, and the model was improved by reducing image feature loss, amplifying global interaction information, fusing crossstage features, and enhancing the feature extraction and generalization ability, which improved the recognition accuracy of five behaviors of laying hens, namely, feeding, drinking, standing, feather arranging, and stooping to search. Meanwhile, the individual identification network was constructed based on the YOLO v8n model, and the individual identification network model was optimized by introducing the MobileNetV3 module, which improved the statistical efficiency of individual behavioral data. The experimental results showed that the optimized behavior identification model achieved 94.4%, 93%, 90.7%, 91.7%, 86.9% average precision (AP) for the recognition of feeding, drinking, standing, feather arranging, and stooping searching behaviors, respectively, and 91.4% mean average precision (mAP), which was comparable to that of YOLO v5n, YOLO v6n, and YOLO v7-tiny, YOLO v8n, the mean average precision mean (mAP) was increased by 4.8, 4.1, 5.5, and 3.5 percentage points, respectively;the number of parameters and the amount of operations of the individual identification model were reduced by 1.9651×106 and 6.1×109 compared with that of the YOLO v8n model.It was found that by analyzing the behavioral data of the laying hens, the behavioral data were related to the temperature and the individual laying hens themselves, and that when the temperature was decreased, the number of feeding and standing was increased, the number of drinking was decreased, the number of finishing feathers and stooping to search almost did not change, the behavioral data of different individual laying hens varied greatly at the same temperature, and the value of the difference was related to the body size of the laying hens. The results of the experiment laid the foundation for judging the health status of laying hens based on behavioral data, precision breeding on farms and preferential selection of individual laying hens.