Soybean diseases are the main factors causing serious yield reduction in soybeans. Therefore, an automatic identification method for soybean leaf diseases is an urgent need to develop. However, most soybean leaf disease recognition models have problems such as weak generalization ability and insufficient labeled soybean images. To overcome these problems, this paper proposes an improved method for unsupervised domain generalization (UDG) named Leaf-Bridge Across Domains (LBADs), which is dedicated to cross-domain soybean leaf disease image recognition. Our method builds a semantic-preserving image-to-image mapping between each training domain and an auxiliary leaf-bridge domain to achieve cross-domain semantic alignment of soybean disease images in the shared feature space through fully unsupervised self-supervised contrastive learning. Firstly, LBAD is performed on both labeled source and target domain data. To bridge the gap between these two domains, it constructs a cross-domain aided leaf-bridge domain within a contrastive self-supervised model, mapping images from different domains into a shared space to achieve semantic alignment. After the fully unsupervised pre-training of the feature backbone, we use 1%, 5%, and 10% of the labeled source domain images, respectively, to fine-tune the linear classifier only (the pre-trained backbone weights are completely frozen) and complete the classification of unlabeled target domain images. The highest accuracy of the three-label fraction can reach 83.56%, 89.22%, and 87.2%, respectively, higher than that of the model without LBAD. The final identification result is also higher than the accuracy of the model trained without LBAD. Experimental results showed that LBAD can improve the generalization ability of the model and obtain a more accurate soybean leaf disease recognition model.
Cui et al. (Fri,) studied this question.