Cross-Domain Leaf Counting with Minimizing Feature Distances

Plant phenotyping refers to quantitative description of the physical and chemical properties of a plant. As an important task for plant phenotyping, leaf count can be used to assess plant growth stages and health. The application of deep learning method in plant phenotype has made important contributions. However, most methods are supervised training, which need to annotate the target dataset for each task, and the cost of annotation is relatively expensive. Therefore, the method of using published or existing annotated datasets to complete cross domain work in the target dataset is worth studying. In this paper, a leaf counting method to minimize the feature distance between the source domain dataset and the target dataset is proposed. After training on the source domain dataset, the counting task can be completed directly on the target dataset. The improved Style-Transfer network module transfer source data to the target domain data, so that the features of the source dataset are close to the features of the target domain dataset. Finally, the density map estimation model is used to train on the generated fake dataset. Experimental results show that the proposed method has good cross-domain counting ability.