基于注意力机制和多尺度残差网络的农作物病害识别

针对传统农作物病害识别方法依靠人工提取特征，步骤复杂且低效，难以实现在田间环境下识别的问题，提出一种多尺度卷积结构与注意力机制结合的农作物病害识别模型。该研究在残差网络（ResNet18）的基础上进行改进，引入Inception模块，利用其多尺度卷积核结构对不同尺度的病害特征进行提取，提高了特征的丰富度。在残差结构的基础上加入注意力机制SE-Net（Squeeze-and-excitation networks），增强了有用特征的权重，减弱了噪声等无用特征的影响，进一步提高特征提取能力并且增强了模型的鲁棒性。实验结果表明，改进后的多尺度注意力残差网络模型(Multi-Scale-SE-ResNet18)在复杂田间环境收集的8种农作物病害数据集上的平均识别准确率达到95.62%，相较于原ResNet18模型准确率提高10.92个百分点，模型占用内存容量仅为44.2MB。改进后的Multi-Scale-SE-ResNet18具有更好的特征提取能力，可以提取到更多的病害特征信息，并且较好地平衡了模型的识别精度与模型复杂度，可为田间环境下农作物病害识别提供参考。

Crop Disease Recognition Based on Attention Mechanism and Multi-scale Residual Network

Aiming at the problem that traditional crop disease recognition methods rely on manual extraction of features, the steps are complex and inefficient, and it is difficult to recognize in the field environment, a crop disease recognition model combining multi-scale convolution structure and attention mechanism was proposed. This research improved on the basis of residual network (ResNet18), introduced the Inception module, and its multi-scale convolution kernel structure was used to extract disease features at different scales, and the richness of features was improved. On the basis of the residual structure, the attention mechanism squeeze-and-excitation networks (SE-Net) was added to enhance the weight of useful features, weaken the influence of useless features such as noise, and further improve the feature extraction ability and enhance the robustness of the model. The experimental results showed that the improved multi-scale attention residual network model (Multi-Scale-SE-ResNet18) had an average recognition accuracy of 95.62% on the eight crop disease data sets collected in a complex field environment, compared with the original accuracy of the ResNet18 model, it was increased by 10.92 percentage points. The model size was only 44.2MB. The improved Multi-Scale-SE-ResNet18 had better feature extraction capabilities, and it can extract more disease feature information, and better balance the recognition accuracy of the model with the model complexity, which can be used for crop diseases identification in the field environment.