Super resolution and image enhancement using novelty concepts

Many remote sensing applications involve operating the sensor in a "staring" mode in which the line-of-sight (LOS) is fixed to a point in the background scene. Inadvertent motions of the line-of-sight either blur the image or create unwanted time signals in the sensor's output. This paper develops a unique technique that actually utilizes the LOS motions between data frames and estimates the underlying scene at a spatial resolution that is "finer" than the basic detector array of the sensor. The approach employs random sampling theory and the concepts of novelty filtering to optimally store all information contained in a set of measured images. The novelty concept develops a set of orthogonal "eigenimages" from a sequence of image frames. The next image is interrogated for "novel" changes not present in the previous frames and the eigenimages are updated with this new information. The LOS motions provide a random sampling of the underlying image. Even though the motions may be random, if they are known then the eigenimages are interpolated to provide an estimate of the detector outputs for any arbitrary sensor pointing. This on be used to either estimate and remove the background from subsequent frames (staring clutter rejection) or provide an image of the underlying background at high spatial resolution (super resolution) The paper presents the theoretical derivation and illustrates the performance of the approach using a set of high fidelity simulated scenes for a space based sensor.