基于残差连接的轻量化番茄作物病害检测模型

Lightweight Tomato Crop Disease Detection Model Based on Residual Connections

摘要:

番茄作物生长过程中容易受到多种病害的侵袭。基于深度学习模型的病害检测策略，其计算量通常较大。针对这个问题，本文提出了一种轻量化深度学习模型ResDepSepNet。该模型基于残差模块构建，引入深度可分离卷积，并通过增加卷积运算步长实现下采样。同时引入SE（Squeeze-and-excitation）注意力模块，提升病害识别能力。采用PlantVillage番茄作物病害数据集对ResDepSepNet模型进行了测试，并将测试结果与MobileNetV2模型和TrioConvTomatoNet模型进行对比。测试结果显示，ResDepSepNet模型的整体识别正确率分别比MobileNetV2模型和TrioConvTomatoNet模型高4.8、1.1个百分点，且其浮点运算次数仅为3.5×107，约为MobilenetV2模型和TrioConvTomatoNet模型的1/18和1/7。本研究为番茄作物病害检测提供了技术参考。

Abstract:

Tomato crops are prone to be attacked by various diseases during their growth process. The computational load of disease detection strategies based on deep learning models was usually substantial. To address this issue, a lightweight deep learning model called ResDepSepNet was proposed. This model was constructed based on residual modules to alleviate the gradient vanishing problem that may occur during model training, thereby improving the overall performance of the model. To reduce the model’s computational load, depthwise separable convolutions were introduced, and downsampling was achieved by increasing the stride of the convolutional operations. Additionally, an squeeze-and-excitation (SE) attention module was introduced to enable the model to focus more on feature information crucial for disease identification, thereby enhancing its disease recognition capability. The ResDepSepNet model was tested by using the PlantVillage tomato disease dataset, and the test results were compared with the MobileNetV2 model and the TrioConvTomatoNet model. The test results showed that the overall accuracy of the ResDepSepNet model was 4.8 and 1.1 percentage points higher than that of the MobileNetV2 and TrioConvTomatoNet models, respectively. Moreover, its floating-point operations count was merely 3.5×107, approximately 1/18 and 1/7 of those of the MobilenetV2 and TrioConvTomatoNet models, respectively. The research result can provide a technical reference for disease detection in tomato crops.