In view of the differences in appearance and the complex backgrounds of crop diseases, automatic identification of field diseases is an extremely challenging topic in smart agriculture. To address this challenge, a popular approach is to design a Deep Convolutional Neural Network (DCNN) model that extracts visual disease features in the images and then identifies the diseases based on the extracted features. This approach performs well under simple background conditions, but has low accuracy and poor robustness under complex backgrounds. In this paper, an end-to-end disease identification model composed of a disease-spot region detector and a disease classifier (YOLOv5s + BiCMT) was proposed. Specifically, the YOLOv5s network was used to detect the disease-spot regions so as to provide a regional attention mechanism to facilitate the disease identification task of the classifier. For the classifier, a Bidirectional Cross-Modal Transformer (BiCMT) model combining the image and text modal information was constructed, which utilizes the correlation and complementarity between the features of the two modalities to achieve the fusion and recognition of disease features. Meanwhile, the problem of inconsistent lengths among different modal data sequences was solved. Eventually, the YOLOv5s + BiCMT model achieved the optimal results on a small dataset. Its Accuracy, Precision, Sensitivity, and Specificity reached 99.23, 97.37, 97.54, and 99.54%, respectively. This paper proves that the bidirectional cross-modal feature fusion by combining disease images and texts is an effective method to identify vegetable diseases in field environments.
Keywords: complex background; cross-modal fusion; disease identification; few-shot; transformer.
Copyright © 2022 Feng, Zhao, Wang, Wu, Miao and Zhang.