Authors

Abstract

Aiming to address the issues of low accuracy and slow response in tomato leaf disease recognition models, an enhanced lightweight model for tomato leaf disease recognition was proposed. The SE attention module in the MobileNetV3-Large model was substituted with a CA attention module, and dilated convolution was incorporated to improve the model's recognition accuracy and response speed. The CA attention module enhances the perception and feature extraction capabilities of spatial coordinate information in images. Dilated convolution was introduced into the deep network architecture to expand the model's receptive field. The model was trained using a transfer learning approach that partially froze specific convolutional layers. Experimental results on a dataset comprising 10 common tomato leaf disease images and healthy leaf images demonstrated that the unimproved model achieved a recognition accuracy of 90.11% and an F1 score of 89.98%. After replacing the SE attention module with the CA attention module, the model's accuracy increased to 91.15%, with the F1 score rising to 91.08%. Furthermore, introducing the dilated convolution model improved the accuracy to 94.33% and the F1 score to 94.22%, while maintaining a parameter count of 2.79×106 and a validation set operation time of 11.76 seconds. Compared to other traditional lightweight models, this model exhibits significant advantages. The DC-CA-MobileNetV3 tomato leaf disease recognition model proposed in this study can accurately and efficiently identify tomato leaf diseases, featuring a small number of parameters and ease of deployment in embedded systems.

Abstract in Chinese

针对番茄叶片病害识别模型的准确度低、响应速度慢的问题，提出了一种改进的轻量级番茄叶片病害识别模型，将MobileNetV3-Large模型中的SE注意力模块替换为CA注意力模块，并引入空洞卷积，以提高模型识别准确度与响应速度。利用CA注意力模块提高对图像空间坐标信息的感知能力和特征提取能力。在深层网络中引入空洞卷积模型，扩大模型感受野。使用只冻结部分卷积层的迁移学习方法对模型进行训练。在常见的10种番茄叶片病害图像与健康叶片图像构成的数据集上的实验结果表明，未改进的模型识别精准率为90.11%，F1值为89.98%；改为CA注意力模块后，模型的精准率提高至91.15%，F1值提高至91.08%；在此基础上引入空洞卷积模型后，精准率提高至94.33%，F1值提高至94.22%，其模型参数量为2.79×106，验证集运行耗时11.76s，与其他传统轻量化模型相对比具有明显优势。本文提出的DC-CA-MobileNetV3番茄叶片病害识别模型可精准、高效地识别番茄叶片病害，同时具有参数量小、易搭载至嵌入式系统的优点。