Satellite Low-Shot Augmented Object Detection (SALSA)

The recent widespread use of overhead sensors, and their ability to provide continuous streams of imagery for intelligence, surveillance and reconnaissance (ISR) missions, has generated a critical need for high-fidelity, automated object detection systems. For intelligence analysts, searching large volumes of imagery with vast spatial and temporal extent can be extremely time consuming and tedious. High-accuracy automated systems could free up a significant portion of an analysts workload, allowing more time for investigative tasks. Objects of interest may not always be commonplace in the data, or known ahead of time, necessitating the ability of such automated systems to be easily trainable with a small amount of training data. We present the Satellite Low-Shot Augmented Detection System (SALSA) to address these challenges. At the core of our approach is a system of deep convolutional neural networks specialized for the low-shot learning of new object categories or subcategories, coupled with advanced generative learning techniques to model how input targets appear in varying conditions.