Advancing Hyperspectral Imaging through IntegratedCompressive Sensing / Inpainting via Machine Learning

We propose to use advanced algorithms for compressive sensing (CS) using inpainting which utilize machine learning to improve the capability of existing or novel hyperspectral imaging (HSI) systems. These approaches will allow for an order of magnitude improvement in the frame acquisition speed of existing detectors at a high compression ratio with no loss of resolution, while at the same time permitting them to overcome object obscuration and radiation damage phenomena in excess of existing technologies in this area. To accomplish these goals we propose to develop new hyperspectral sensors that intentionally omit pixels that will be recovered using model based compressive sensing. This provides the ability to selectively acquire a small number of pixels in an existing imaging array to form the complete image with no loss of image resolution or fidelity (or the ability to select an even smaller number with an acceptable loss of resolution). In addition, the algorithms that will be developed can adapt to the form of key image types using machine learning, so that the number of pixels can be reduced further, or that the pixels used for the image can be adaptively changed to avoid obscuration/damaged pixels. This type of imaging system has significant benefits for imaging/detection across a whole range of NASA programs and missions for recovery of damaged equipment or scenes, image enhancement, data compression or using novel designs with reduced pixel densities. This imaging approach will also automatically compress all images obtained by the level of sub-sampling used to form the image, reducing transmission and storage costs significantly, even without using additional lossless data compression before transmission.