| 摘 要: 针对小麦病虫害分类识别研究中卷积神经网络加深导致的梯度消失和爆炸等问题,提出了一种基于改进GoogLeNet的识别方法。首先,构建小麦病虫害数据集 Wheat11,并通过数据增强技术保持样本间的平衡;其次,引入CBAM注意力机制,并改进注意力模块,增强网络对病虫害特征的提取能力;最后,在Inception模块中引入残差结构,有效缓解梯度消失和爆炸等问题。试验结果表明,改进模型在测试集上的准确率达到94.50%,比未改进前
提升3.30%。该模型在小麦病虫害上有更好的分类识别效果,为后续小麦病虫害防治提供方法指导。 |
| 关键词: 小麦病虫害 GoogLeNet 注意力机制 残差结构 卷积神经网络 |
|
中图分类号: TP391
文献标识码: A
|
|
| Research on Classification of Wheat Diseases and Pests Based on Improved GoogLeNet |
|
WANG Jianfeng, LI Yue
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
a1393607499@163.com; liyue@gsau.edu.cn
|
| Abstract: To address issues such as gradient vanishing and explosion caused by deepening convolutional neural networks in wheat disease and pest classification research, this study proposes an improved GoogLeNe-t based
recognition method. Firstly, the Wheat11 dataset (with 11 categories) was constructed, with sample balance maintained through data augmentation techniques. Secondly, the CBAM attention mechanism was introduced, and the attention
module was optimized to enhance the network’s feature extraction capability for disease and pest characteristics.Finally, residual structures were integrated into the Inception modules, effectively mitigating gradient vanishing and
explosion problems. Experimental results demonstrate that the improved model achieved an accuracy of 94.50% on the test set, representing a 3.30-percentage-point improvement compared with the original model. This model exhibits
superior classification and recognition performance for wheat diseases and pests, providing methodological guidance for subsequent prevention and control efforts. |
| Keywords: wheat diseases and pests GoogLeNet attention mechanism residual structure convolutional neural network |
