| 摘 要: 针对传统深度学习网络模型因权重参数过大而不适宜在移动端部署的问题,提出了一种基于改进的轻量级MobileNetV3网络模型对苹果叶片病害进行识别。利用PyTorch框架,选取26 377张图片,按6∶2∶2的比例划分数据集,以MobileNetV3网络模型为基础,引入迁移学习,加入空洞卷积,并调整原模型结构,形成新模型进行训练学习。经过多轮迭代,损失曲线实现收敛,模型训练成功,训练集上的准确率为95.72%,测试集上的准确率为93.41%;经过改进的MobileNetV3网络模型对苹果叶片病害图像识别效果较好,为将来在移动端实现部署与推广提供了技术方案支撑。 |
| 关键词: 神经网络;MobileNetV3;深度学习;迁移学习;空洞卷积 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 甘肃省科技计划项目(20CX9NA095) |
|
| Research on Image Recognition Technology for Apple Leaf Diseases Based on Improved MobileNetV3 |
|
WANG Zhibing, WU Lili
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
1904235213@qq.com; wull@gsau.edu.cn
|
| Abstract: Aiming at the problem that traditional deep learning network model is not suitable for deployment on mobile terminal due to too large weight parameters, this paper proposes an improved lightweight MobileNetV3 network model to identify apple leaf diseases. PyTorch framework is used to select 26 377 pictures and divide the data set according to 6:2:2. Based on the MobileNetV3 network model, transfer learning is introduced, dilated convolution is added, and the structure of the original model is adjusted to form a new model for training and learning. After several iterations, the loss curve converges and the model is successfully trained, with an accuracy of 95.72% on the training set and 93.41% on the test set. The improved MobileNetV3 network model performs well in identifying apple leaf disease images, providing technical support for future deployment and promotion on mobile terminals in future. |
| Keywords: neural networks; MobileNetV3; deep learning; transfer learning; dilated convolution |
