With the `real' data defined, the image is scanned for prominent sources (using the median-filtering NIPS algorithm for speed). The approximate radius and centroid position of such sources are then used to establish source contaminated pixels that are to be ignored when evaluating the background model.
The image is then binned into boxes of user defined size (but modified if necessary to fit the image) and the mean in each box is determined, by fitting a Poisson distribution function to the brightness histogram of the pixels in the box (in an attempt to bias against remaining weak undetected sources). An uncertainty on each box mean is also computed. A weighted smoothing function (top-hat or Gaussian of user defined size - but modified to accommodate holes left by ignored sources) is then applied to the resulting grid of box mean values.
The background surface thus computed can be (optionally) subtracted from the input image. At the same time, statistics about the quality of the match between the background model and the input image are derived and output.