Signature Research, Inc. (SGR) and Michigan Technological University (MTU) propose a Phase I STTR effort to develop a learning algorithm which exploits the spatio-spectral characteristics inherent within IR imagery and motion imagery.Our archive of modelled and labeled data sets will allow our team to thoroughly capture the variable elements that will drive machine learning performance.The overall effort will determine the feasibility of using a learning algorithm to augment the GEOINT professionals analysis of IR look-down order of battle analysis.The specific objectives of the effort include establishing what makes an IR image suitable for use in terms of machine learning recognition performance and determining whether the application of machine learning data processing toward IR image recognition is appropriate for further testing.It is critical to have a diverse training data set for any machine learning algorithm to exploit and process IR imagery due to the wide variety of target environments.A robust algorithm using a cross-validation approach for data-derived model learning will be created by augmenting existing IR imagery sets with simulated thermal scenes, physics-based rendering, atmospheric, and empirical sensor models, allowing the algorithm to exploit targets in a wide variety of conditions.
