Usually the software image form of FilterImage is be converted to another form; e.g. ImageBuffer -> PixelBuffer. But we rarely need both forms of image. There are instances of filters where we fail to render the filtered image just because we cache more than form of the image of FilterImage. See the test case in bug 242219.

So we should keep only one form of the image. And we can get the original form of the image from the new one if we want. This will reduce the memory needed to generate the software filtered image.