基于全局特征提取的农作物病害识别模型

针对现阶段特征提取网络当测试样本出现歪斜、模糊、缺损等变化时识别效果不够理想，利用训练样本扩充、变换、缩放等方式改善网络性能并不能动态地满足实际的复杂病害图像识别任务的问题，在ResNet50中引入双层注意力机制与通道特征提取机制，设计基于全局特征提取的深度学习网络（Global feature deep learning network，GFDL-Net），该网络包括通道特征提取子网络（Squeeze and excitation net，SE-Net）和双注意力特征提取子网络（Double feature extraction net，DFE-Net），分别从通道空间特征提取与平面关键点特征提取两方面改善了网络的全局特征提取能力。为了验证GFDL-Net的有效性，对辣椒、马铃薯、番茄等15种病害图像加入不同角度的旋转、色彩变换等测试，发现在样本加入旋转后与ResNet50、BoTNet、EfficientNet相比，平均识别准确率分别高出20.05、18.62、21.97个百分点；加入明暗度、饱和度、对比度变换后与ResNet50、BoTNet、 EfficientNet相比，平均识别准确率分别高出3.57、0.53、3.98个百分点，而识别速度分别为ResNet50、BoTNet、EfficientNet的4.4、4.9、2.0倍。试验证明GFDL-Net在图像全局特征提取能力方面的改进能有效提升网络的泛化能力与鲁棒性，可将其应用于解决变化样本的农作物病害识别任务中。

Deep Learning Network for Crop Disease Recognition with Global Feature Extraction

In view of the fact that the recognition effect of the feature extraction network at this stage is not ideal when the test samples are skewed, fuzzy, defective and other changes, improving the network performance by expanding, transforming, scaling and other ways of training samples cannot dynamically meet the problem of the actual complex disease image recognition task, In ResNet50, a global feature deep learning network (GFDL-Net) based on global feature extraction was designed by introducing a two-layer attention mechanism and channel feature extraction mechanism. The network included channel feature extraction sub network (Squeeze and exception net, SE-Net) and double feature extraction net (DFE-Net), the global feature extraction ability of the network was improved from two aspects: channel space feature extraction and plane key point feature extraction. In order to verify the effectiveness of GFDL-Net, tests such as rotation at different angles and color transformation were added to the images of 15 diseases such as pepper, potato and tomato. It was found that the average recognition accuracy was 20.05 percentage points, 18.62 percentage points and 21.97 percentage points higher than that of ResNet50, BoTNet and EfficientNet respectively after adding rotation to the samples. Compared with ResNet50, BoTNet and EfficientNet, the average recognition accuracy was 3.57 percentage points, 0.53 percentage points and 3.98 percentage points higher, and the recognition speed was 4.4 times, 4.9 times and 2.0 times of ResNet50, BoTNet and EfficientNet respectively after adding the shading, saturation and contrast transformations. The experiment proved that the improvement of GFDL-Net in the global feature extraction ability of images can effectively improve the generalization ability and robustness of the network, which can be used to solve the crop disease recognition task of changing samples.