GFNet: Gnostic Fields based Low-Shot Learning for Target Detection in Remote Sensing

To detect uncommon targets in remote sensing imagery, it is quite often that very few prior examples are available. This so-called low-shot detection remains a very challenging problem in remote sensing, despite the recent development in state-of-the-art object detection algorithms such as Faster R-CNN and YOLO, and low-shot learning methods such as feature shrinking, model regression and memory augmented neural networks. To achieve human-level performance in low-shot learning on remote sensing images, a deep learning architecture that more closely resembles human vision learning is worthy of further study. Motivated by the human vision systems natural low shot learning capability, Intelligent Automation, Inc. (IAI), along with our collaborator Dr. Christopher Kanan from Rochester Institute of Technology (RIT), proposes to develop a novel deep neural network based target detection system called GFNet, that excels in low-shot learning in remote sensing imagery. The key innovation of the proposed approach is a brain-inspired classifier called gnostic field that can be viewed as a type of semi-parametric neural network.