| 摘要: |
| 为提高设施生产中对各生长阶段生菜鲜重的无损估测精度进而更好地指导生产，提出一种利用生菜冠层图像为输入，基于迁移学习技术和卷积神经网络估测鲜重的方法，对比分析AlexNet、VGG-16、GoogLeNet和ResNet-18模型迁移学习后在生菜鲜重估测任务上的效果；同时，对比不同迁移学习方法对模型性能的影响，通过冻结卷积层和减少全连接层改善模型的参数量和训练速度。结果表明:1)AlexNet和VGG-16两种模型能较好的实现生菜鲜重的估测，AlexNet模型的决定系数R2为0.928 0，标准均方根误差NRMSE为19.08%，VGG-16模型的R2为0.938 0，NRMSE为17.71%，但VGG-16模型存在参数量大训练慢的问题，综合考虑选取AlexNet模型迁移学习后作为生菜鲜重估测模型；2)与全新学习方法相比，在预训练模型基础上对生菜鲜重数据集进行迁移学习，可以明显提升生菜鲜重估测模型的训练速度和准确度；3)冻结卷积层能显著加快模型的训练速度，训练时间可减少18%，减少全连接层在保持精度的前提下能大幅度减少模型的参数量。基于迁移学习的卷积神经网络模型可用于生菜鲜重的快速估测，该方法也可以拓展应用到其他叶类蔬菜的鲜重估测中。 |
| 关键词: 生菜 鲜重估测 卷积神经网络 迁移学习 |
| DOI：10.11841/j.issn.1007-4333.2022.11.17 |
| 投稿时间：2022-02-07 |
| 基金项目:国家重点研发计划战略性科技创新合作专项(2020YFE0203600)；基本科研业务费(Y2021PT04) |
|
| Lettuce fresh weight estimation based on transfer learning and convolutional neural network |
|
CUI Tingyuan1,YANG Qichang2,ZHANG Yi1,XU Dan3,MA Juncheng1
|
| (1.Institute of Environment and Sustainable Development in Agricultural, Chinese Academy ofAgricultural Sciences, Beijing 100083, China;2.Institute of Urban Agriculture, Chinese Academy of Agricultural Sciences, Chengdu 610000, China;3.College of Water Resources and Civil Engineering, China Agricultural University, Beijing 100083, China) |
| Abstract: |
| To improve the accuracy of non-destructive estimation of lettuce fresh weight at different growth stages in facility production and to better guide production, based on transfer learning technology and convolution neural network, a new estimating method by using canopy image as input was proposed. The results were compared and analyzed after transfer learning on AlexNet, VGG-16, GoogLeNet, and ResNet-18 models. At the same time, the effects of using different transfer learning methods on the performance of the model were also compared. The training speed and parameters of the model were improved by freezing the convolution layers and reducing the number of full connection layers. The results showed that: 1)AlexNet and VGG-16 models were better to estimate the fresh weight. The determination coefficient R2 of the AlexNet model was 0. 928 0, the standard root mean square error NRMSE was 19. 08%, while the R2 of the VGG-16 transfer learning model was 0. 938 0, and the NRMSE was 17. 71%, respectively. However, the VGG-16 model had the problems that it needed a large number of parameters and was slow to be trained. Therefore, the AlexNet model was selected as the fresh weight estimation model after transfer learning; 2)Compared with the new learning methods, transfer learning based on the pre-trained model could significantly improve the training speed and accuracy; 3)Freezing the convolutional layers could significantly accelerate the training speed of the model and reduce the training time by 18% and reducing the fully connected layers could significantly reduce the number of parameters of the AlexNet model. In summary, the convolution neural network model based on transfer learning can be used for faster estimation of the fresh weight for lettuce, and the method can also be extended and applied to the fresh weight estimation of other leafy vegetables. |
| Key words: lettuce fresh weight estimation convolutional neural network transfer learning |
