Small Sample Meta-leaming Towards Object Recognition Through UAV Observations

Deep learning has achieved impressive results for image classification and object recognition. However, deep learning typically requires too many labeled data for model training which is not always applicable in the tasks that data collection is difficult or expensive. In this paper, we consider the problem of object recognition by a UAV that can observe a given object from different viewpoints and learn to recognize similar objects. Specifically, we have used the Recursive Cortical Network (RCN) for car recognition based on a recorded UAV video. Furthermore, we have combined meta-learning and transfer learning to further improve the training efficiency. Experiments results have demonstrated the effectiveness of the proposed method using the Ominiglot dataset.