Résumé
Classification of plants based on a multi-organ approach is very challenging due to the variability in shape and appearance in plant organs. Despite promising solutions built using convolutional neural network (CNN) for plant classification, the existing approaches do not consider the correspondence between different views captured of a plant. In fact, botanists usually observe and study simultaneously a plant from different vintage points, as a whole and also analyse different organs in order to disambiguate species. Driven by this insight, we introduce a new framework for plant structural learning using the recurrent neural network (RNN) approach. This novel approach supports classification based on a varying number of plant views composed of one or more organs of a plant, by optimizing the dependencies between them. We also present the qualitative results of our proposed models by visualizing the learned attention maps. To our knowledge, this is the first study to venture into such dependencies modeling and interpret the respective neural net for plant classification. Finally, we show that our proposed method outperforms the conventional CNN approach on the PlantClef2015 benchmark. The source code and models are available at https://github.com/cschan/Deep-Plant.