基于改进MobileNetV3的水稻病害识别模型

针对水稻病害识别方法准确度低、模型收敛速度缓慢的问题，本文提出了一种高性能的轻量级水稻病害识别模型，简称为CA（Coordinate attention）-MobileNetV3。通过微调的迁移学习策略完善了模型的训练，提升了模型收敛速度。首先创建10个种类的数据集，其中包含9种水稻病害和1种水稻健康叶片。其次使用CA模块，在通道注意力中嵌入空间坐标信息，提高模型的特征提取能力与泛化能力。最后，将改进后的MobileNetV3网络作为特征提取网络，并加入SVM多分类器，提高模型精度。实验结果表明，在本文构建的水稻病害数据集上，初始的MobileNetV3识别准确率仅为95.78%，F1值为95.36%；加入CA模块后识别准确率和F1值分别提高至96.73%和96.56%；再加入SVM多分类器，通过迁移学习后，改进模型的识别准确率和F1值分别达到97.12%和97.04%，参数量和耗时仅为2.99×106和0.91s，明显优于其他模型。本文提出的CA-MobileNetV3水稻病害识别模型能够有效识别水稻叶部病害，实现了轻量级、高性能、易部署的水稻病害分类识别算法。

Rice Disease Identification Model Based on Improved MobileNetV3

For the problems of low accuracy of rice disease identification methods and slow convergence of models, a high performance lightweight rice disease identification model was proposed, referred to as coordinate attention (CA)-MobileNetV3. The training of the model was optimized by fine-tuning the migration learning strategy, and the convergence speed of the model was improved. Firstly, a ten species dataset was created, containing nine rice diseases and healthy rice leaves. Secondly, the CA module was also used to embed spatial coordinate information in the channel attention to improve the feature extraction and generalization ability of the model. In addition, the improved MobileNetV3 network was used as the feature extraction network and the SVM multi-classifier was added to improve the model accuracy. The experimental results showed that on the rice disease dataset constructed, the initial MobileNetV3 recognition accuracy was only 95.78% and the F1 score was 95.36%, and then the recognition accuracy and F1 score were improved to 96.73% and 96.56%, respectively, after adding the CA module, and then the SVM multi classifier was added, and the recognition accuracy and F1 scores reached 97.12% and 97.04%, respectively, the number of parameters and the time taken were only 2.99×106 and 0.91s, which were significantly better than that of other models. The experimental results showed that the CA-MobileNetV3 rice disease recognition model proposed can effectively recognize rice leaf diseases and achieve a lightweight, high-performance and easy-to-deploy rice disease classification and recognition algorithm.