Abstract: In order to address the issues of numerous network parameters, large model memory and long training time in recognizing diseases and pests on pepper leaves in natural environments by using the GoogLeNet model, a lightweight GoogLeNet model incorporating the CBAM mechanism (CBAM—GoogLeNet) is proposed. In this model, Inception (4b) and Inception (4c) modules are replaced by the CBAM attention mechanism, while this attention mechanism is inserted into  the average pooling layer, and L2 regularization is added in the fully connected layer, so as to reduce the model size and shorten the training time, while ensuring high accuracy and validation rates of the network model. A visual pepper disease and pest identification system is also designed by using the MATLAB platform. The results show that the size of the CBAM—GoogLeNet model is reduced by 91.2%, 96.2%, 96.3%, and 15.0% compared to AlexNet, VGG16, VGG19, and GoogLeNet, respectively. The training time is reduced by 12.7%, 26.5%, 62.2%, and 8.8%, respectively. Additionally, the model achieves an identification accuracy of 99.5% and a validation accuracy of 97.3%, realizing the goals of model lightweighting and fast, accurate recognition. This provides effective technical support for timely prevention and control of pepper diseases and pests, and the reduction of losses.

Key words:  , pepper pests and diseases, accurate recognition, lightweight model, attention mechanism, deep learning

摘要： 针对GoogLeNet模型在自然环境下进行辣椒叶片病虫害识别时存在网络参数多、模型内存大以及训练时间长的问题，提出一种融合CBAM机制的轻量化GoogLeNet模型(CBAM—GoogLeNet)。采用CBAM注意力机制替换Inception(4b)和Inception(4c)模块，将该注意力机制插入到平均池化层之后，在全连接层中添加L2正则化，达到减小训练模型和缩短训练时长的目的，同时保证网络模型的高准确率和验证率，并结合MATLAB平台设计一款可视化的辣椒病虫害识别系统。结果表明，CBAM—GoogLeNet的模型大小相比AlexNet、VGG16、VGG19和GoogLeNet分别缩小91.2%、96.2%、96.3%和15.0%，训练时长分别减少12.7%、26.5%、62.2%和8.8%，此外，该模型的识别准确率达到99.5%，验证准确率达到97.3%，实现模型轻量化和快速精准识别的目标。为辣椒及时防治、减少损失提供一种有效的技术支持。

关键词: 辣椒病虫害, 精准识别, 轻量化模型, 注意力机制, 深度学习