The emergence of machine learning into scientific fields has created opportunities for novel and powerful image processing techniques. Algorithms that can perform complex tasks without a “man-in-the-loop” or explicit instructions are invaluable artificial intelligence tools. These algorithms typically require a large set of training data on which to base statistical predictions. In the case of electro-optical infrared (EO/IR) remote sensing, algorithm designers often seek a substantial library of images comprising many weather conditions, times of day, sensor resolutions, etc. These images may be synthetic (predicted) or measured, but should encompass a large variety of targets imaged from a variety of vantage points against numerous backgrounds. Acquiring such a large set of measured imagery with sufficient variation can be difficult, requiring numerous field campaigns. Alternatively, accurate prediction of target signatures in cluttered outdoor scenes may be a viable option. In this work, sensor imagery is generated using CoTherm, a co-simulation tool which operates MuSES (an EO/IR simulation code) in an automated fashion to create a large library of synthetic images. The relevant MuSES inputs – which might include environmental factors, global location, date and time, vehicle engine state, human clothing and activity level, or sensor waveband – can be manipulated by a CoTherm workflow process. The output of this process is a large library of MuSES-generated EO/IR sensor radiance images suitable for algorithm development. If desired, synthetic target pixels can be inserted into measured background images for added realism.
|