国家自然科学基金(31870532); 湖南省自然科学基金(2021JJ31163); 湖南省教育科学“十三五”规划基金(XJK20BGD048)
目的 水稻产量关乎全人类的粮食安全，如何有效地预防和高效地检测水稻病虫害是智慧农业领域的重要课题。深度学习由于具备自主学习图像特征等优异性能，成为水稻病虫害识别的首选方法。但在自然环境下，数据集偏小，且容易受到复杂背景的影响，在训练过程中容易产生过拟合，以及细微特征难以提取等问题。本研究致力于解决上述问题。方法 提出一种基于改进ResNet的多尺度双分支结构的水稻病虫害识别模型(MSDB-ResNet)。在ResNet模型的基础上，引入ConvNeXt残差块，以优化残差块的计算比例，构建双分支结构，通过调整每条分支的卷积核大小，提取输入病害图像中大小不同的病害特征。针对现实环境复杂、数据集太小、过拟合等问题，利用从自然环境拍摄到的5932张水稻病虫害图像，使用随机亮度、运动模糊等数据预处理方法，以及镜像、裁剪、缩放等数据增强方法，将数据集扩充到20000张，训练MSDB-ResNet模型识别4种常见的水稻病害。结果 MSDB-ResNet在水稻病虫害数据集上具有良好的识别性能，识别准确率高达99.10%，较原ResNet 模型提高了2.42个百分点，明显优于AlexNet、VGG、DenseNet和ResNet等经典网络。该模型具有良好的泛化能力和极强的鲁棒性。结论 MSDB-ResNet模型在水稻病虫害识别中具有一定的可行性和先进性，可为实现复杂背景下的水稻病虫害识别提供参考。
Objective The yield of rice is related to food security of all mankind. How to effectively prevent and detect rice diseases and pests is an important topic in the field of smart agriculture. Deep learning has become the preferred method for identifying rice diseases and pests due to its excellent performance in self-learning image features. However, in natural environments, the dataset is relatively small and susceptible to complex backgrounds, resulting in overfitting and difficulty in extracting subtle features during training. This study aims to address the aforementioned issues. Method We proposed a rice disease and pest identification model with multi-scale dual branch structure based on improved ResNet (MSDB ResNet). On the basis of the ResNet model, ConvNeXt residual blocks were introduced to optimize the calculation proportion of residual blocks, construct a dual branch structure, and extract disease features of different sizes from the input disease image by adjusting the convolution kernel size of each branch. In response to issues such as complex real world environments, small datasets, and overfitting, a total of 5932 rice pest and disease images captured from natural environments was utilized. Using data preprocessing methods such as random brightness and motion blur, as well as data augmentation methods such as mirroring, cropping, and scaling, the dataset was expanded to 20000 pictures. The MSDB-ResNet model was trained to identify four common rice diseases. Result MSDB-ResNet had good recognition performance on rice disease and pest datasets, with a recognition accuracy of 99.10%, which was 2.42 percentage points higher than the original ResNet model and obviously superior to classic networks such as AlexNet, VGG, DenseNet, ResNet, etc. This model had good generalization ability and strong robustness. Conclusion The MSDB ResNet model is feasible and progressiveness in the identification of rice diseases and pests, which provides a reference for the identification of rice diseases and pests under complex background.