Known and unknown class recognition on plant species and diseases

Recognizing plant species and disease is essential to practical applications, such as keeping biodiversity and obtaining a desired crop yield. This study aims to extend the recognition from known to unknown classes in the context of plants, termed Plant-relevant Open-Set Recognition (POSR). In this task, a trained model is required to either classify an input image into one of the known classes or an unknown class, even if the model is only trained with the images of known classes. To achieve this task, we propose a method to obtain a high-performance classifier with compact feature distributions for known classes. To have a high-performance classifier, a ViT model pre-trained in the PlantCLEF2022 dataset is transferred, following an observation that a plant-related source dataset is more beneficial to plant species and disease recognition than other commonly used datasets, such as ImageNet. To have compact feature distributions, we adopt additive margin Softmax loss (AM-Softmax) which brings the distance smaller between the features of the same known class and hence gives more spaces for the unknown class. Extensive experimental results suggest that our method outperforms current algorithms. To be more specific, our method obtains AUROC 93.685 and OSCR 93.256 on average on four public datasets, with an average accuracy of 99.295 on closed-set classification. We believe that our study will contribute to the community and, to fuel the field, our codes will be public2.