轻量级多场景群养猪只行为识别模型研究

Research of Lightweight Multi-scene Group Pig Behavior Recognition Model

针对现有猪只行为识别模型体积大、识别场景单一、部署应用硬件要求高等问题,本文提出轻量级多场景群养猪只行为识别模型YOLO v5n-PBR(YOLO v5n for pig behavior recognition)。首先通过拍摄和收集不同养殖场景、不同猪只数量及不同角度的群养猪只行为数据构建多场景群养猪只行为数据集,并根据该数据集中猪只行为目标的特点引入迁移学习方法和OTA(Optimal transport assignment)标签分配方法对YOLO v5n模型进行训练,加快模型收敛速度并提升模型精度,构建高精度多场景群养猪只行为识别模型;然后利用L1-norm剪枝算法筛选并删减模型中不重要的通道,去除冗余参数;最后通过微调训练和中间特征知识蒸馏去除剪枝带来的性能劣化,从而得到轻量级多场景群养猪只行为识别模型YOLO v5n-PBR并进行嵌入式设备部署。实验结果表明,YOLO v5n-PBR模型平均精度均值(mean average precision,mAP)为96.9%,参数量、计算量和内存占用量分别为4.700×10^(5)、1.20×10^(9)和1.2 MB,在两种不同系统和不同硬件配置的嵌入式设备上的部署实时识别帧率分别为12.2帧/s和66.3帧/s,与原始模型YOLO v5n相比,mAP提高1.1个百分点,参数量、计算量和内存占用量分别减少73.3%、70.7%和68.4%,部署实时识别帧率分别提高74.3%和83.1%。此外,基于多场景群养猪只行为数据集训练得到的YOLO v5n-PBR模型在4个单场景或双场景的群养猪只行为数据集上的mAP均能达到98.1%,对2种不同养殖场景的6段猪只行为视频的嵌入式设备部署识别统计结果与人工统计结果相近,平均精确率和平均召回率均为95.3%,以较少的参数达到较强的泛化性。本文提出的YOLO v5n-PBR模型具有精度高、体积小、速度快、泛化性强等优点,可满足嵌入式设备部署要求,为猪只行为的实时、准确监测及猪只行为识别模型的部署应用提供技术基础。

In order to solve the problems of large size,single recognition scene and high hardware requirements for deploying application of existing pig behavior recognition models,a lightweight multi-scene group pig behavior recognition model YOLO v5n for pig behavior recognition(YOLO v5n-PBR)was proposed.Firstly,a multi-scene group pig behavior dataset was constructed by shooting and collecting group pig behavior data from different breeding scenes,different pig numbers and different angles,and based on the characteristics of pig behavior objectives in the dataset,the transfer learning method and the optimal transport assignment label assignment method were introduced to train the YOLO v5n model,which accelerated the model convergence speed and improved the model accuracy,and a high-precision multi-scene group pig behavior recognition model was constructed.Then the L1-norm pruning algorithm was used to screen and delete the unimportant channels in the model to remove the redundant parameters.Finally,the performance degradation caused by pruning was removed by fine-tuning training and intermediate feature knowledge distillation,so that the lightweight multi-scene group pig behavior recognition model YOLO v5n-PBR was obtained and deployed as embedded devices.Experimental results showed that the mean average precision(mAP)of the YOLO v5n-PBR model was 96.9%,with parameters,amount of computation,and memory footprint being 4.700×10^(5),1.20×10^(9),and 1.2 MB,respectively.The deploy real-time recognition frame rates on embedded devices with different systems and hardware configurations were 12.2 frames/s and 66.3 frames/s.Compared with that of the original YOLO v5n model,the mAP was improved by 1.1 percentage points,and parameters,amount of computation,and memory footprint were decreased by 73.3%,70.7%,and 68.4%,respectively.The deploy real-time recognition frame rates were increased by 74.3%and 83.1%.In addition,the YOLO v5n-PBR model trained based on the multi-scene group pig behavior dataset can reach 98.1%of mAP on four single-scene or dual-scene group pig behavior datasets,and the statistical results of embedded device deployment recognition of six pig behavior videos in two different breeding scenes were similar to those of manual statistics,with an average accuracy and average recall rate of 95.3%,which achieved strong generalization with fewer parameters.The YOLO v5n-PBR model proposed had the advantages of high accuracy,small size,fast speed,and strong generalization,which can meet the deployment requirements of embedded devices and provide a technical basis for real-time and accurate monitoring of pig behavior and the deploying application of pig behavior recognition model.