Analysis of Infrared Data (AID) via Transfer Learning on Deep Networks

In recent years deep learning techniques and other machine learning methods applied to Full Motion Video (FMV) have re-defined the state-of-the-art in terms of ability to detect and classify objects of interest in imagery and video. This recent success has been made based upon extremely large data sets being available for training. The Army has a need to leverage these object classification capabilities in imaging domains relevant to the Army. Unfortunately, the DoD is typically lacking in the amount of data (including data of military significant target types) to properly implement modern learning techniques such as deep learning. The deficiency of data is primarily apparent in the thermal infrared (IR) domain. The Army is seeking techniques that can exploit highly trained and effective learning models from large visible and civilian datasets (or artificially constructed datasets) by transferring the models to perform detection and analysis on similar but smaller military significant IR and other data. The technique is generally called transfer learning. Ideally the transfer learning would enable rapid adjustment of trained IR models to new target types and environments on the fly. This program will support the Army Modernization Priority: Next Generation Combat Vehicle (NGCV).