Synthetic Image Generation for AI/ML

The rapid development of deep learning has enabled powerful image recognition algorithms. These algorithms require a large set of training images on which to base statistical predictions, and often the acquisition of a suitable image library (in terms of quality and quantity) is a limiting factor. The problem of finding measured imagery is especially significant outside of the visible waveband. Imagery of specific targets in the near-infrared (NIR), short-wave infrared (SWIR) and ōthermalö wavebands (e.g., mid-wave infrared and long-wave infrared, referred to as MWIR and LWIR, respectively) are nearly impossible to find in the public domain. Consequently, synthetic imagery can be an appropriate alternative content source. With simulated scenes, the simulation engineer has control over the aforementioned variables, resulting in the ability to create robust algorithms and gauge the impact of variables independently. In this Phase 2 proposal, we detail an efficient procedure for the creation of a large library of synthetic electro optic/infrared (EO/IR) images suitable for automated target recognition (ATR) algorithm development and testing. The goal of this project is to refine a process for building synthetic IR image datasets that successfully train computer vision and ATR algorithms, and then transfer that process to the Army. This will be accomplished through AI/ML algorithm training and testing, process automation for synthetic image generation, and expanding a database of IR target models to include cluttered urban and natural scenes with civilian and soldier dismounts with a variety of clothing and poses. The MuSES and CoTherm software programs will also be modified to improve accuracy and ease of use. ThermoAnalytics is the developer of MuSES, the physics-based EO/IR signature prediction software. MuSES, and its commercial counterpart TAITherm, have been in active development for 25 years through USG support and significant IRAD investment. We plan to leverage the image generation capabilities of MuSES to create sets of synthetic training data with variations representing our investigation space.