Orchid classification using homogeneous ensemble of small deep convolutional neural network

Orchids are flowering plants in the large and diverse family Orchidaceae. Orchid flowers may share similar visual characteristics even they are from different species. Thus, classifying orchid species from images is a hugely challenging task. Motivated by the inadequacy of the current state-of-the-art general-purpose image classification methods in differentiating subtle differences between orchid flower images, we propose a hybrid model architecture to better classify the orchid species from images. The model architecture is composed of three parts: the global prediction network (GPN), the local prediction network (LPN), and the ensemble neural network (ENN). The GPN predicts the orchid species by global features of orchid flowers. The LPN looks into local features such as the organs of orchid plant via a spatial transformer network. Finally, the ENN fuses the intermediate predictions from the GPN and the LPN modules and produces the final prediction. All modules are implemented based on a robust convolutional neural network with transfer learning methodology from notable existing models. Due to the interplay between the modules, we also guidelined the training steps necessary for achieving higher predictive performance. The classification results based on an extensive in-house Orchids-52 dataset demonstrated the superiority of the proposed method compared to the state of the art.