关键词: 苹果病害, 图像识别, 生成对抗网络, 残差网络, 混合注意力机制

Abstract: It is important to identify and control the disease accurately to improve the yield and quality of apples. Aiming at the problem of low recognition accuracy of apple disease images caused by small lesion areas and easy background interference, a Convolutional Block Attention ModuleResidual Network （CBAM-ResNet） model was designed. This model embedded a tandem channel attention module and spatial attention module in the residual network so that the model can focus on the diseased areas of image features and improve the recognition accuracy. Aiming at the problem that the number of samples in the model training dataset is insufficient and imbalanced, this paper proposed a method for dataset expansion based on Wasserstein Generative Adversarial Networks （WGAN）. Through the antagonistic training between generator and discriminator, 10 000 apple disease images were generated, and CBAM-RESNET was trained and tested to improve the generalization ability and robustness of the model. The experimental results showed that compared with traditional models such as CNN, VGG-16, ResNet-50, and Inception-V3, CBAM-ResNet had a better recognition effect on apple diseases, and its recognition accuracy, precision, recall, and F1-Score reached 95.50%, 95.40%, 95.40%, and 0.95. The method can provide technical support for accurate identification and realtime monitoring of apple disease images.

Key words: apple disease, image recognition, generative adversarial networks, residual network, convolutional block attention module