改进RegNet识别多种农作物病害受害程度

针对传统农作物病害识别方法效率低、受害程度识别不准确的问题，提出了一个基于深度迁移学习和改进RegNet的多种农作物病害受害程度识别模型。该模型首先在RegNet输入端进行在线数据增强，用以提高训练样本的多样性；其次在模型的特征提取层引入了有效通道注意力机制，用以提高模型的特征提取能力；然后在模型的分类层引入多尺度特征融合策略，用以提高模型对细粒度特征的分类能力；最后使用深度迁移学习来优化模型的整体性能，加快模型的收敛速度，提高模型的泛化能力。试验结果表明，改进后的网络模型在农作物病害受害程度数据集上准确率达到了94.5%，相较于RegNet原模型准确率提高了10.4个百分点。改进后的模型具有更好的特征提取能力，对细粒度特征有更强的分类能力，该模型为农作物病害类型及其受害程度的识别提供了新的方法。

Identifying the damage degree of various crop diseases using an improved RegNet

Crop yield has been one of the most prominent issues in the world in recent years. However, crop diseases have posed a great threat to crop yield. It is a high demand to timely and accurately detect crop disease types and the degree of damage. The manual recognition can rely only on skilled technicians. But, the visual fatigue of humans can easily lead to reduce the accuracy rate. The current machine learning cannot consider the correlation between the attributes in the data set, resulting in low recognition accuracy. In this study, a network model was proposed to identify the damage degree of multiple crop diseases using deep transfer learning and improved RegNet. The model contained four aspects as follows. Firstly, an online data enhancement was carried out at the input side of this model. Nine strategies were selected for the data enhancement, such as the HSV color variation, grayscale transformation, and Gaussian noise. The diversity of data samples increased while reducing the time and space for the data set collection expansion. As such, the over-fitting of the network was alleviated during this time. Secondly, the Ef?cient Channel Attention (ECA) mechanism was introduced into the feature extraction layer of the model for the cross-channel interaction. The model was then improved to extract more subtle features, particularly for the crop disease features. As such, a higher accuracy of recognition was achieved for the crop disease damage, which increased by 1.5 percentage points, compared with the original model at the same model size. The F1 score also increased by 2.2 percentage points, compared with the previous one. In addition, a multi-scale feature fusion was introduced into the classification layer of the model. A spatial pyramid pooling was adopted to highly improve the accuracy of the model. The degree of crop disease damage was classified at different scales, especially for the fine-grained features. Correspondingly, the accuracy of crop disease damage recognition increased by 2.4 percentage points, and the F1 score increased by 1.5 percentage points more than before. Finally, deep transfer learning was used to optimize the overall performance of the model. The convergence speed was accelerated to improve the generalization ability of the model. The recognition accuracy was improved by 2.7 percentage points, compared with the strategy without deep transfer learning. The experimental results show that the improved RegNet network model achieved 94.5 percentage points accuracy on the dataset of crop disease damage level, which was 10.4 percent higher than the original one. The recognition accuracy of the improved model was improved by 2.1, 6, 3.7, and 1.6 percentage points, respectively, while the model sizes were reduced by 35.8, 458.1, 35.4, and 52.1MB, respectively, compared with the commonly-used classification network models, such as ResNet50, VGG16, InceptionV3, and ConvNeXt. Consequently, higher accuracy of recognition and smaller model size were achieved in the improved classification model, compared with the rest. The better performance of feature extraction and stronger classification ability were also obtained for the fine-grained features during this time. The finding can also provide a promising way to identify the crop disease types and the degree of damage.