HA-Net: A High-Accuracy Network for Rice Diseases Recognition

Rice plants are susceptibility to pest infestations and various diseases. Convolutional neutral networks (CNNs) have greatly improved accuracy and speed for rice diseases. However, important textures are lost with the depth of CNNs increases and simple classifier potentially lead to the results getting trapped in a local optimum. To address the above issues, we design a high-accuracy network (HA-Net) for rice disease diagnosis in the early stage. The HA-Net consists of multi-encoding fusion module and average classification module. The multi-encoding fusion module integrates the advantages of different VGG-16 and ResNet-50 backbone to explore potential semantics by highlighting the differences of two networks. The proposed average classification module significantly improve accuracy by averaging the results of MLP and SVM. Extensive experiments are conducted on the public RLD dataset and the results demonstrate the superiority of the proposed HA-Net.