On this effort Toyon Research Corp. and The Pennsylvania State University are developing deep learning-based algorithms for object recognition and new class discovery in look-down infrared (IR) imagery. Our approach involves the development of a hybrid classifier that exploits both transfer learning and semi-supervised paradigms in order to maintain good generalization accuracy, especially when limited labeled examples but potentially many unlabeled data exist. Furthermore, the classifier will be able to discover new object classes and target signatures not found in the training data but are well-suited for IR data exploitation. We will also develop algorithms for the generation of infrared images of a given class of interest from one modality (for which available data resources may be scarce) from images from another modality (for which available data resources may be plentiful). This method exploits the paradigm of deterministic annealing to learn associations between pairs of images from the two modalities available image databases.
