This SBIR Phase I project proposes a deep learning approach to the reliable detection and fine-grained classification of a specific model of aircraft in high-resolution unclassified commercial satellite imagery. If adopted, this automated visual surveillance would reduce the current time requirements of the existing manual process by dozens of man-hours. The system takes advantage of recent advances in computer vision technologies and machine learning to identify discriminative visual features that indicate the presence of these aircraft in satellite images. The system will be developed to most effectively utilize existing massive unlabelled satellite imagery datasets to increase performance metrics while reducing image annotation and development costs. The Phase I effort will include: development of automated aircraft detection capabilities, development of an aircraft classifier that can identify aircraft as being of a specific type, quantitative and qualitative evaluation of the technologies, demonstration of the proof of concept using real-world data in a dashboard interface, and a feasibility study to establish the effectiveness of self-supervised learning techniques in reducing development costs by utilizing unlabelled satellite imagery datasets.
