基于多尺度特征提取深度残差网络的水稻害虫识别

【目的】在水稻生产过程中，针对不同虫害需要采用不同的防治方案，水稻害虫的准确识别分类是制定针对性防治方案的前提。【方法】采用深度学习结合机器视觉的方法，基于Res2Net结构提出了一种多尺度特征提取的深度残差网络，通过准确地提取害虫特征实现复杂自然背景下的水稻害虫识别；采用改进的残差结构，使用等级制的类残差连接取代了原本的3×3卷积核，增加了每个网络层的感受野，可以更细粒度地提取多尺度特征。【结果】本网络训练的模型能够有效地识别自然背景下的水稻害虫，在自建的包含22类常见水稻害虫的图像数据集上，平均识别准确率达到了92.023%，优于传统的ResNet、VGG等网络。【结论】本文提出的模型可应用于水稻虫情自动监测系统，为实现水稻害虫虫情的机器视觉监测提供参考。

Rice pest recognition based on multi-scale feature extraction depth residual network

Objective In the process of rice production, different control schemes need to be adopted for different pests. The accurate identification and classification of rice pests are the premise of formulating targeted control program. Method A deep residual network of multi-scale feature extraction was proposed based on the Res2Net structure, which could extract pest characteristics more accurately and realize rice pest identification in complex natural background. This network adopted an improved residual structure, replaced the original convolutional kernel with hierarchical class residual connections, increased the sensing field of each network layer, and could extract multi-scale features at a more fine-grained degree. Result The results showed that the model trained by this network could effectively identify rice pests in natural background. The average recognition accuracy of proposed model reached 92.023% on the self-built image dataset containing 22 kinds of the common rice pests, which was superior to the traditional ResNet, VGG and other networks. Conclusion This network can be applied to the automatic monitoring system of rice insect status, which provides a reference for the realization of machine vision monitoring of rice pests.