In view of the obvious differences in the manifestations of the same diseases in apples at different stages of the disease, different diseases show certain similarities, and the early symptoms of the disease are not obvious. For these problems, a new model attention residual network (ARNet) was introduced based on the combination of attention and residual thought. The model introduces the multi-layer attention modules to solve the problems of early disease location dispersion and features that are difficult to extract. In order to avoid network degradation, a residual module was constructed to effectively integrate high and low-level features, and data augment technology was introduced to prevent the model from over-fitting. The proposed model (ARNet) achieved an average accuracy of 99.49% on the test set of 4 kinds of apple leaf diseases with real complex backgrounds. Compared with the models ResNet50 (99.19%) and MobileNetV2 (98.17%), it had better classification performance. The model proposed in this paper had strong robustness and high stability and can provide a reference for the intelligent diagnosis of apple leaf diseases in practical applications.

The key to diagnosing the types and degree of apple leaf diseases is to correctly segment apple leaf disease spots. Therefore, in order to effectively solve the problem of poor segmentation of leaves and diseased areas, the U2Net semantic segmentation network model was used in the research of apple leaf disease identification and disease diagnosis, and compared with the classic semantic segmentation network model DeepLabV3+ and UNet. In addition, the effects of different learning rates (0.01, 0.001, 0.0001) and optimizers (Adam, SGD) on the performance of U2Net network model were compared and analyzed. The experimental results showed that the learning rate is 0.001 and the optimizer is Adam, the average pixel accuracy (MPA) and mean intersection over union (MIoU) of the research model for lesion segmentation reach 98.87% and 84.43%, respectively. The results of this study were expected to provide the theoretical basis for the precise control of apple leaf disease.

Because of the high similarity of leaves of different Cerasus humilis varieties, it is difficult to identify them with the naked eye. In this study, the leaves of four different Cerasus humilis varieties collected in the field were used as the research objects, and a new leaf recognition model based on the improved lightweight convolution neural network model EfficientNet-B0 was proposed. Firstly, the performance of the network models Efficientnet-B0 and ResNet50, GoogleNet, ShuffleNet, and MobileNetV3 were compared based on two different learning methods. Then, the influence of different optimizers on model recognition accuracy was compared based on the optimal model. Finally, different learning rates were used to optimize the optimal model. The results show that the recognition rate of the proposed Efficientnet-B0 +Ranger+0.0005 model was up to 86.9%, which was 2.23% higher than that of the original Efficientnet-B0 model. The results show that this method can effectively improve the recognition accuracy of Cerasus humilis auriculate leaves, which can provide a reference for the deployment of the leaf identification model of Cerasus humilis variety on the mobile terminal.