| 基于迁移学习的苹果落叶病识别与应用 |
| |
| 引用本文： | 郭惠萍,曹亚州,王晨思,荣麟瑞,李怡,王霆伟,杨福增.基于迁移学习的苹果落叶病识别与应用[J].农业工程学报,2024,40(3):184-192. |
| |
| 作者姓名： | 郭惠萍 曹亚州 王晨思 荣麟瑞 李怡 王霆伟 杨福增 |
| |
| 作者单位： | 西北农林科技大学机械与电子工程学院，杨凌 712100;西北农林科技大学机械与电子工程学院，杨凌 712100;西安比亚迪汽车工程研究院，西安 710000 |
| |
| 基金项目： | 陕西省科技重大专项（2020zdzx03-04-01）；陕西省科技重大专项（2020zdzx03-04-02）；陕西省技术创新引导专项（2021QFY08-05） |
| |
| 摘 要： | 为解决现有卷积神经网络苹果叶片病害识别模型泛化能力弱，模型体积较大等问题，该研究提出一种基于改进MobileNetV3苹果落叶病识别模型。以健康叶片和常见苹果落叶病为研究对象，包括斑点落叶病、灰斑病、褐斑病、锈病4种，每种病害2级，共9类特征，通过改进网络的注意力模块、全连接层及算子，结合迁移学习的训练方式，构建苹果落叶病识别模型。在扩充前后的数据集上对比不同的学习方式、学习率和注意力模块等对模型的影响，验证模型的识别性能。试验结果表明：采用迁移学习的方式，在训练50轮达曲线收敛，比全新学习的准确率增加6.74～10.79个百分点；使用引入的ET（efficient channel attention-tanh）注意力模块，网络损失曲线更加平滑，模型的参数量更少，模型体积减小了48%，提高了模型的泛化能力；在扩充数据集上，学习率为0.000 1时，结合迁移学习的训练方式，改进MobileNetV3（ET3-MobileNetV3）苹果落叶病识别模型，平均准确率能达到95.62%，模型体积6.29 MB。将模型部署到喷药设备上，可实现基于苹果叶片病害识别的变量喷施，该研究可为苹果叶片病害的检测与果园的现代化管理提供参考。
|
| 关 键 词： | 病害 图像识别 苹果落叶病 ET注意力模块 改进MobileNetV3 迁移学习 |
| 收稿时间： | 2023/8/4 0:00:00 |
| 修稿时间： | 2024/1/5 0:00:00 |
Recognition and application of apple defoliation disease based on transfer learning |
| |
| GUO Huiping,CAO Yazhou,WANG Chensi,RONG Linrui,LI Yi,WANG Tingwei,YANG Fuzeng.Recognition and application of apple defoliation disease based on transfer learning[J].Transactions of the Chinese Society of Agricultural Engineering,2024,40(3):184-192. |
| |
| Authors: | GUO Huiping CAO Yazhou WANG Chensi RONG Linrui LI Yi WANG Tingwei YANG Fuzeng |
| |
| Abstract: | Convolutional Neural Network (CNN) can be applied to recognize the leaf disease of apples in agricultural production, due to the reduced size of the model and the high generalization. In this study, the recognition model was proposed for the apple defoliation disease using an improved MobileNetV3 (ET3-MobileNetV3) network structure. According to the disease features, the images were divided into five types: altermaria boltch, rust, grey spot, brown spot and health. The images of apple defoliation disease were collected from the standard dataset produced by the Luochuan Apple Experimental Station of Northwest A&F University. The orchard dataset was collected from Huicheng Apple Farm in Yangling City. The final datasets of apple defoliation diseases were a total of 21 950 images, including 19 819 in the training set and 2 131 in the test set. The recognition model of defoliation disease was constructed to improve the attention module. Efficient Channel Attention (ECA) was replaced by Squeeze and Excite (SE), while the Tanh function was used to replace the Sigmoid function. Then, the full connection layer was improved to replace the original Hard-Swish (HS) activation function with the ReLU6 activation function. At the same time, the Dropout layer and the improved Bottleneck operator were introduced to enhance the calculation speed. Finally, transfer learning was utilized to transfer the pre-trained weights into the recognition task in the training. Among them, the pre-trained weights were obtained to train the model on the ImageNet dataset. The generalization was improved on the small sample learning, whereas, there was a decrease in the data demand of the target task, indicating the better performance of the model. The performance of recognition was verified by training the model on datasets before and after expansion, transfer learning and new learning, different learning rates and attention mechanisms. The learning rates included 0.01, 0.001, and 0.000 1. Attention modules included SE, ECA and Efficient Channel Attention-Tanh (ET). The experimental results showed that the transfer learning training obtained a superior performance model in a shorter time, compared with the fully new learning. The model curve converged faster than before, and the more stable curve was achieved after convergence. The generalization was stronger, and the recognition accuracy was higher than before. When the learning rate of transfer learning was 0.000 1, the model curve tended to converge after 50 rounds of training. There was a lower standard deviation in the accuracy and Loss curves, where the accuracy increased by 6.47 to 10.79 percentage points. Secondly, the ET module was introduced using the fully connected layer. The convergence of the loss curve was accelerated to reduce the number of parameters. The generalization was improved to avoid the overfitting of the model. The final model volume was 6.29M, which was reduced by 48%. The standard deviation of the Loss curve was 0.006 4, indicating a smoother Loss curve. Thirdly, the diversity and quantity of data increased to expand the datasets, thereby reducing the overfitting of the model. At the same time, there was an increase in the diversity of data exposure to model training. The better generalization was suitable for the complex data. The prediction accuracy was also improved for the robustness and stability of the model. The average accuracy and F1-score reached 95.62% and 94.62, respectively, under transfer learning combined with data expansion. Finally, the ET3-MobileNetV3 model was deployed to the spraying equipment. The apple defoliation diseases were effectively identified with the variable spraying. The high accuracy, strong generalization, and small parameter volume fully met the requirements of disease recognition during spraying in the orchard. The optimal model was deployed in the orchard spraying device. The pesticide spraying was also realized to implement the guidance variable spraying device. The finding can provide a strong reference to detect apple defoliation diseases in modern orchards. |
| |
| Keywords: | disease image recognition apple defoliation disease ET attention module improved MobileNetV3 transfer learning |
|
| 点击此处可从《农业工程学报》浏览原始摘要信息 |
| 点击此处可从《农业工程学报》下载免费的PDF全文 |
|
