Multi-View Fusion Network for Crop Disease Recognition.

During the growth of crops, crop yields will be affected by various diseases. Automatic and accurate recognition of crop diseases and determination of disease severity are the key to crop disease prevention and control. In practice, due to noise interference in the data collection process, etc., the acquired crop disease recognition datasets are often with large intra-class differences and large inter-class similarities, which brings great challenges to the development of accurate crop disease recognition. Existing crop disease recognition methods often ignore the subtle differences in easily confused categories, which leads to limited performance. To this end, this paper proposes a multi-view fusion network (MVF-Net) for crop disease recognition. The proposed MVF-Net consists of two parallel branches: suppression branch and global branch. The suppression branch explores discriminative fine-grained features by suppressing the saliency information of the image. The global branch aims to learn the global features of the image to avoid the network paying too much attention to local features. In addition, in order to promote network learning with more discriminative features, this paper introduces an embedding loss, which is combined with the traditional multi-class cross-entropy loss function to improve the performance of the crop disease recognition model. Extensive experiments are conducted on the 2018ai_challenger crop disease recognition dataset. The results show that the proposed MVF-Net has certain advantages over other comparison methods, which lays the foundation for the prevention and control of crop diseases.