首页|基于MobileNetV3的植物叶片识别系统

基于MobileNetV3的植物叶片识别系统

扫码查看
针对现有植物叶片识别研究存在的模型泛化性差的问题,本文设计一个基于MobileNetV3-Large网络和迁移学习的植物叶片识别系统.通过自采集图像补充数据和使用图像锐化、翻转、亮度增强等方法构建了包含32种植物的大规模叶片数据集,以MobileNetV3-Large网络和预训练权重为前提,寻找最佳超参数完成模型的迁移学习,对32种植物叶片进行特征提取和分类.通过PyQt5的前后端部署,该方法被实例化为一个实用性强的植物叶片识别系统.在测试集上的实验结果表明,MobileNetV3-Large 达到 98.45%的识别准确率,与 AlexNet、ResNet 和 MobileNetV2相比分别提高12.46%、1.09%和9.62%,有效弥补了模型泛化性差的短板.该系统对32类植物的叶片的识别效果颇佳,满足各种场景下的植物叶片种类识别的需求.
Plant leaf recognition system based on MobileNetV3
Aiming at the problem of poor model generalization in the existing plant leaf recognition research,this paper designs a plant leaf recognition system based on MobileNetV3-Large network and transfer learning.A large-scale leaf data set containing 32 plants is constructed by means of supplementary data of self-collected images and using image sharpening,flipping,and brightness enhancement.Based on the MobileNetV3-Large network and pre-training weights,the optimal hyperparameters are found to complete the transfer learning of the model,and the feature extraction and classification of 32 plant leaves are performed.Through the front-end and back-end deployment of PyQt5,this method is instantiated as a practical plant leaf recognition system.The experimental results on the test set demonstrates that MobileNetV3-Large achieves a recognition accuracy of 98.45%,which is 12.46%,1.09%and 9.62%higher than that of AlexNet,ResNet and MobileNetV2,respectively,effectively making up for the shortcomings of poor generalization of the model.The system has a good recognition effect on the leaves of 32 kinds of plants,and meets the needs of plant leaf species recognition in various scenarios.

lightweight convolutional neural networkMobileNetV3-Largetransfer learningsystem designplant leaf recognition

张柔绮、赵家松、严伟榆

展开 >

云南农业大学理学院,云南昆明 650201

福州大学物理与信息工程学院,福建福州 350108

云南农业大学大数据学院,云南昆明 650201

轻量级卷积神经网络 MobileNetV3-Large 迁移学习 系统设计 植物叶片识别

云南省农业基础研究联合专项基金云南省教育厅科学研究基金

202301BD070001-2022019J0171

2024

云南民族大学学报(自然科学版)
云南民族大学

云南民族大学学报(自然科学版)

CSTPCD
影响因子:0.381
ISSN:1672-8513
年,卷(期):2024.33(4)