Apple Leaf Disease Regcognition Method Base on Improved ShuffleNet V2

Most of the apple leaf disease images acquired in natural scenes contain complex backgrounds. At the same time, the disease features are more likely to appear at arbitrary locations in the images due to the less stringent shooting requirements. All these factors affect the recognition accuracy of convolutional neural networks. To address this problem, a lightweight apple leaf disease recognition method based on improved ShuffleNet V2 is proposed. Methods based on ShuffleNet V2 network model, max-pooling was introduced to remove the interference information such as complex background to a certain extent, on the other hand, to reduce the excessive sensitivity of the model to feature location. Then, BAM(bottlenect attention module) was integrated to enable the model to learn more important disease features and reduce the interference of information such as complex background. Experiments on apple leaf disease dataset show that the improved model has 1.3M network parameters, 145.7MFLOPs of computation, and 1.5% recognition error rate, which is 2% lower compared to ShuffleNet V2. The proposed method has the advantages of low recognition error rate, lightweight model, and applicability to apple leaf disease images acquired in natural scenes, so it has stronger practicability.