摘要
【目的】探究残差神经网络(residual neural network,ResNet)对不同种类鸡蛋的分类效果,明确深度学习应用存在智能鸡蛋巡检装置的可行性,为家禽养殖智能化进程提供新思路,并为鸡蛋分类研究提供数据支撑。【方法】在鸡舍实地取样,采用自适应矩估计优化器(adaptive moment estimation,Adam)以微调最后1层、微调所有层和重新训练所有层3种迁移学习策略分别训练,并通过调整模型权重参数及改变学习率的方式训练出最佳分类模型。【结果】得到识别准确率高达98.971%的鸡蛋分类模型。计算出模型在数据集上的各类评估指标,并借助混淆矩阵及语义特征降维可视化,分析出鸡蛋分类识别中易被误判的类别及语义。该模型部署后实时性良好,满足实际需求。【结论】鸡蛋的分类识别中光照条件是关键影响因素,应尽可能使鸡舍光照稳定均衡。针对6类鸡蛋,微调所有层并调整学习率参数为0.6,可得最佳模型。其在鸡舍场景下分类效果优良,尤其是颜色语义,应用于智能鸡蛋巡检装置,可有效降低人力成本。后续研究中应注重畸形蛋及软壳蛋的记录,为进一步优化提供数据支撑。
【Objective】This research was conducted to explore the classification performance of residual neural networks(ResNet)on different types of eggs,clarify the feasibility of applying deep learning in intelligent egg inspection devices,provide new ideas for the intelligent process of poultry farming,and provide data support for egg classification research.【Method】Field sampling was conducted in the chicken coop,and an adaptive moment estimation optimizer(Adam)was used to train three transfer learning strategies:fine-tuning the last layer,fine-tuning all layers,and retraining all layers.The optimal classification model was trained by adjusting the model weight parameters and changing the learning rate.【Result】An egg classification model was obtained with a recognition accuracy of up to 98.971%.The various evaluation indicators of the model on the dataset was calculated,and confusion matrix and semantic feature dimensionality reduction visualization was used to analyze the categories and semantics that are prone to misjudgment in egg classification recognition.The model has good real-time performance after deployment and can meet practical needs.【Conclusion】The lighting conditions are a key influencing factor in the classification and recognition of eggs,and the lighting in the chicken coop should be kept stable and balanced as much as possible.For six types of eggs,the best model can be obtained by fine-tuning all layers and adjusting the learning rate parameter to 0.6.It has excellent classification performance in chicken coop scenes,especially in color semantics.When applied to intelligent egg inspection devices,it can effectively reduce labor costs.In subsequent research,attention should be paid to recording deformed eggs and soft shell eggs to provide data support for further optimization.
作者
梁旭
王玲
赵书涵
LIANG Xu;WANG Ling;ZHAO Shuhan(College of Mechanical and Electrical Engineering,Henan Agricultural University,Zhengzhou 450002,China)
出处
《河南农业大学学报》
CAS
CSCD
北大核心
2024年第3期456-466,共11页
Journal of Henan Agricultural University
基金
河南省科技攻关项目(232102321022,232102110284)
河南省高等学校青年骨干教师培养计划(2020GGJS046)。
关键词
鸡蛋分类
家禽养殖
残差神经网络
学习率
智慧农业
迁移学习
egg classification
poultry farming
residual neural network
learning rate
smart agriculture
transfer learning
