基于轻量化网络与注意力机制的育肥猪采食行为识别方法研究  

Study on feeding behavior recognition method of fattening pigsbased on lightweight networks and attention mechanism

作　　者：陆舟 沈明霞[1] 刘龙申[1] 薛鸿翔 丁奇安 陈佳 LU Zhou;SHEN Mingxia;LIU Longshen;XUE Hongxiang;DING Qi’an;CHEN Jia(College of Engineering/Jiangsu Intelligent Animal Husbandry Equipment Science and Technology Innovation Center,Nanjing Agricultural University,Nanjing 210031,China)

机构地区：[1]南京农业大学工学院/江苏智慧牧业装备科技创新中心,江苏南京210031

出　　处：《南京农业大学学报》2023年第4期802-812,共11页Journal of Nanjing Agricultural University

基　　金：江苏省科技计划项目(BE2019382)。

摘　　要：[目的]针对育肥猪采食行为识别误差大、检测速度慢等问题,提出一种具有轻量化结构的育肥猪采食行为检测模型,实现对育肥猪采食行为的快速检测与采食时长统计。[方法]以YOLO v5L目标检测算法为基础,构建侧视视角下的猪只采食行为检测模型。对比更换不同轻量化主干网络后对模型检测效果的影响,选取性能最优的模型;改进ShuffleNet V2网络结构基本单元,采用Mish激活函数提高模型泛化能力与推理速度,引入SE注意力机制给予目标特征更高的权重以提高目标识别精度;对比分析模型增加非营养性访问行为检测前、后的采食行为识别准确率。[结果]优化后的育肥猪采食行为检测模型大小为38.2 MB,计算量为37.8 GFLOPs,视频检测平均帧耗时7.6 ms。与非营养性访问行为进行区分识别后,猪只采食行为检测识别准确率为96.4%,召回率为92.5%。模型检测采食时长与人工统计采食时长相对误差为6.1%。[结论]改进的YOLO v5L-ShuffleNet网络模型检测速度和模型大小均能满足实际生产需求,可在复杂养殖环境中全天候识别育肥猪采食行为。[Objectives]Aiming at the problems of large identification error and slow detection speed of feeding behavior of fattening pigs,a feeding behavior detection model of fattening pigs with lightweight structure was proposed,which realized the rapid detection of feeding behavior of fattening pigs and the statistics of feeding duration.[Methods]Based on the YOLO v5L object detection algorithm,a pig feeding behavior detection model from a side-view perspective was constructed.Compared with the impact of replacing different lightweight backbone networks on the detection effect of the model,the model with the best performance was selected.The basic unit of ShuffleNet V2 network structure was improved,and the group convolution and channel mixing operations were used to reduce the network calculation amount and fully integrated the characteristic information between different channels.The Mish activation function was used to improve the generalization ability and inference speed of the model,and the SE(squeeze and excitation)attention mechanism was introduced to give higher weight to the target features to improve the target recognition accuracy.The model was compared and analyzed to increase the accuracy of feeding behavior recognition before and after the detection of non-nutritive visit behavior.[Results]The size of the optimized feeding behavior detection model of the fattening pig was 38.2 MB,the calculation amount was 37.8 GFLOPs,and the average frame time of video detection was 7.6 ms.After increasing the non-nutritive visit behavior detection,the accuracy of the detection of feeding behavior in pigs in this model was 96.4%,and the recall rate was 92.5%,respectively.The relative error between the model’s detection of feeding time and artificial statistical feeding time was 6.1%.[Conclusions]The improved YOLO v5L-ShuffleNet network model’s detection speed and the model size can meet the actual production requirements,and can identify pig feeding behavior and record feeding time around the clock in complex breedin

关 键 词：育肥猪 采食行为 轻量化模型 YOLO v5L 实时监测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S828[自动化与计算机技术—计算机科学与技术]