摘要
针对荷斯坦奶牛个体识别神经网络的人工调参成本高、泛化性差、效率低,难以实现局部遮挡条件下精准识别等问题,提出了一种基于ResNet框架和神经网络架构搜索(NAS)的自适应网络参数优化算法(NAS-Res)。首先,通过设计包含CBR_K1、CBR_K3、CBR_K5和SkipConnect的操作集,配合密集连接路径,构成超参数网络。然后基于梯度下降的搜索策略,在多目标优化复合损失函数的约束下,强化了对低成本模型的设计。结果表明,NAS-Res在GPU上仅耗时6.18 h获得最佳架构,在包含168头奶牛局部遮挡侧面图像的PO-Cows数据集上,闭集验证准确率为90.18%,与ResNet-18、ResNet-34和ResNet-50相比提高5.04、3.02、14.92个百分点,而参数量分别降低5.9×10^(5)、1.069×10^(7)和1.317×10^(7)。在包含174头奶牛背部图像的Cows2021数据集上闭集验证准确率为99.25%。此外,NAS-Res可忽略PO-Cows数据集规模变化的影响,牛只数量在50~168头之间变化时,Top-1准确率和Top-5准确率变化幅度仅为1.51、1.01个百分点,适用性较强。总体而言,NAS-Res算法实现了对局部遮挡奶牛的精准个体识别,本研究可为复杂背景下畜禽个体识别提供技术参考。
The Holstein cow individual recognition network has the problems of high parameter adjustment cost,poor generalization and low efficiency,and it is difficult to achieve accurate recognition under partial occlusion conditions.An adaptive network parameter optimization identification algorithm(NAS-Res)was proposed based on ResNet framework and neural network architecture search(NAS).Firstly,a hyperparameter network was constructed by designing an operation set,including CBR_K1,CBR_K3,CBR_K5,and SkipConnect,together with dense connection paths.Then the search strategy based on gradient descent strengthened the design of a low-cost model under the constraint of multiobjective optimization composite loss function.The results showed that NAS-Res only took 6.18 GPU hours to obtain the best architecture.On the PO-Cows dataset,which contained side images of 168 cows,NAS-Res achieved 90.18%Top-1 Acc.Compared with ResNet-18,ResNet-34,and ResNet-50,the accuracy was improved by 5.04 percentage points,3.02 percentage points,and 14.92 percentage points,respectively,while the parameters were reduced by 5.9×10^(5),1.069×10^(7),and 1.317×10^(7),respectively.It achieved 99.25%accuracy on the Cows2021 dataset,which contained 174 back images of cows.In addition,NAS-Res can ignore the influence of the scale change of the PO-Cows dataset,and when the number of cattle was changed between 50 and 168,the change range of Top-1 Acc and Top-5 Acc was only 1.51 percentage points and 1.01 percentage points,which showed strong applicability.In general,the NAS-Res algorithm achieved accurate individual identification of partially occluded cows,and the research result can provide technical reference for individual identification of livestock and poultry under complex background.
作者
姚冲
李前
刘刚
吕树盛
侯冲
张淼
YAO Chong;LI Qian;LIU Gang;LÜ Shusheng;HOU Chong;ZHANG Miao(Key Laboratory of Smart Agriculture Systems Integration,Ministry of Education,China Agricultural University,Beijing 100083,China;Key Laboratory of Agricultural Information Acquisition Technology,Ministry of Agriculture and Rural Affairs,China Agricultural University,Beijing 100083,China;Machinery Industry Technology Service Center,Chinese Academy of Agricultural Mechanization Sciences Group Co.,Ltd.,Beijing 100083,China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2023年第S01期252-259,共8页
Transactions of the Chinese Society for Agricultural Machinery
基金
国家自然科学基金项目(32272930)
关键词
荷斯坦奶牛
局部遮挡
神经网络架构搜索
卷积神经网络
个体识别
Holstein cows
partially occluded
neural architecture search
convolutional neural network
individual recognition
