Linear filtering of spatially invariant image sequences for feature separation and compression under three types of image noise

Many medical imaging, remote sensing and other important imaging applications result in a set of images called a linearly additive, spatially invariant (LA SI) image sequence. Previous research has shown that a K-image sequence with M distinct features (M<K) can be linearly filtered using either the KL transform or the orthogonal projection (OP) transform into an M-image set from. which the original K-image set can be recovered with complete feature reconstruction. In those studies, the noise was modeled as Gaussian (uniform over the image scene) or Poisson (dependent on the feature distribution). This paper demonstrates the effectiveness of the same technique for LA SI image sequences with salt and pepper noise, which is neither uniform nor feature dependent, by comparing the results with those for the same image scene with an equal amount of Gaussian or Poisson noise