Hi Warren,
Georg has explained everything important here. DBE is mostly about image statistics. It implements a robust rejection algorithm that knows how to gather background pixels avoiding outlier structures such as stars, hot and dead pixels, cosmic rays, etc.
However, as always happens when we are working with real world data, we have to deal with uncertainty. In this case, uncertainty arises when deciding if a given pixel belongs to the background or not. One of the most important lessons learned after some time involved in signal processing is that 'yes' and 'no' are both bad answers, in general, to this kind of questions. For this reason I implemented something more sophisticated in DBE: fuzzy logic. So we don't ask whether a pixel is part of the background or not; instead, we try to figure out how is it likely for a given pixel to be a background pixel. For each DBE sample, you can find an estimate of its probability of being representative of the true background of the image (whatever this happens to be) as the Wr, Wg and Wb values, respectively for the red, green and blue channels, where applicable.
Of course, this isn't a perfect system: in difficult cases (really wild gradients, complex-shaped gradients, dense star fields, etc.) we cannot remove uncertainty so easily, and then some manual verification work is required. Each particular gradient removal case tends to be a unique challenge. Feel free to upload one of the images where you're having trouble with DBE and we'll be glad to give it a try here, time permitting.
