基于MobileNetV3Small-ECA的水稻病害轻量级识别研究

为实现水稻病害的轻量化识别与检测，使用ECA注意力机制改进MobileNetV3Small模型，并使用共享参数迁移学习对水稻病害进行智能化轻量级识别和检测。在PlantVillage数据集上进行预训练，将预训练得到的共享参数迁移到对水稻病害识别模型上微调优化。在开源水稻病害数据集上进行试验测试，试验结果表明，在非迁移学习下，识别准确率达到97.47%，在迁移学习下识别准确率达到99.92%，同时参数量减少26.69%。其次，通过Grad-CAM进行可视化，本文方法与其他注意力机制CBAM和SENET相比，ECA模块生成的结果与图像中病斑的位置和颜色更加一致，表明网络可以更好地聚焦水稻病害的特征，并且通过可视化和各水稻病害分析了误分类原因。本文方法实现了水稻病害识别模型的轻量化，使其能够在移动设备等资源受限的场景中部署，达到快速、高效、便携的目的。同时开发了基于Android的水稻病害识别系统，方便于在边缘端进行水稻病害识别分析。

Lightweight Identification of Rice Diseases Based on Improved ECA and MobileNetV3Small

In order to realize the lightweight identification and detection of rice diseases, the ECA attention mechanism was used to improve the MobileNetV3Small model, and shared parameter transfer learning was used to carry out intelligent lightweight identification and detection of rice diseases. Pre-training was performed on the PlantVillage dataset, and the shared parameters obtained from the pre-training were transferred to the rice disease recognition model for fine-tuning and optimization. Experiments were on the open-source rice disease dataset. The experimental results showed that the recognition accuracy rate reached 97.47% under non-transfer learning, and 99.92% under transfer learning, while reducing the number of parameters by 26.69%. Secondly, the Grad-CAM was used for visualization. Compared with other attention mechanisms CBAM and SENET, the results generated by the ECA module were more consistent with the position and color of the disease spots in the image, indicating that the network can better focus on rice diseases. Characteristics, and the causes of misclassification were analyzed through visualization and each rice disease. The proposed method realized the lightweight of the rice disease recognition model, so that it can be deployed in resource-constrained scenarios such as mobile devices, and achieved the purpose of fast, efficient and portable. At the same time, an Android-based rice disease identification system was developed, which can facilitate the identification and analysis of rice diseases at the edge.