Authors

Abstract

The precise detection and recognition are the premise in accurate prevention and control of tomato diseases. To improve the accuracy of tomato diseases recognition model, nine kinds of sick leaves images including tomato target spot bacteria in Plant Village and healthy leaves images were used. A new attention mechanism module called CBAM-Ⅱ was created by changing the serial connection between Channel and Spatial attentions of CBAM to parallel connection, and then the results of two modules were added together. CBAM-Ⅱ had been verified to be effective and universal in the convolutional neural network model. The accuracy of MobileNet-V2 with CBAM-Ⅱ model was 99.47%，which had increased by 1.13%, 0.93%, 0.7%8 and 1.06 % respectively comparing with MobileNet-V2 model, MobileNet-V2 plus Channel attention module, MobileNet-V2 plus Spatial attention module, and CBAM attention module. Furthermore, the accuracy of AlexNet, Inception-V3 and ResNet50 model has increased 1.73, 0.15 and 0.33 % respectively when the CBAM-Ⅱ module was added. Results showed that the proposed module CBAM-Ⅱ created in this experiment is more effective in MobileNet-V2 model for tomato diseases recognition, and could solve interference problems resulted from the serial connection. Additionally, the accuracy of four convolutional neural network models including Mobilenet-V2, AlexNet, Inception-V3 and ResNet50 model had all increased when the CBAM-Ⅱ module was added, which represented the good universality of CBAM-Ⅱ module. The results could provide technical support in accurate detection and control of tomato diseases.

Abstract in Chinese

番茄病害准确检测与识别是番茄病害精准化防治的前提。为提高番茄病害识别模型准确率，本文以Plant Village中番茄斑点病等9类病害叶片及健康叶片图像为研究对象，将CBAM的Channel attention和Spatial attention模块由串行连接变为并行连接并把运算结果相加，提出了一种新的注意力机制模块CBAM-Ⅱ，并验证该模块在卷积神经网络模型中的有效性和通用性。在MobileNet-V2模型加入CBAM-Ⅱ模块后，模型准确率达到了99.47%，准确率较MobileNet-V2原模型、加入Channel attention模块、Spatial attention模块以及CBAM注意力模块的MobileNet-V2模型分别提升了1.13、0.93、0.78、1.06个百分点；在AlexNet、Inception-V3和ResNet50加入CBAM-Ⅱ模块后，识别准确率较原模型分别提升了1.73、0.15和0.33个百分点。研究表明，提出的CBAM-Ⅱ在番茄病害识别模型MobileNet-V2中更有效，且可解决CBAM在番茄病害识别过程中串行连接而引起的干扰问题；加入CBAM-Ⅱ的MobileNet-V2、AlexNet等4种卷积神经网络模型较原模型的准确率均有提高，表明通用性好。本文研究可为番茄病害精准检测和防治提供技术支持。