Lets start by looking at the influence of this filtering on smooth color gradients. To test the impact, I have created a test image with a smooth color gradient going from red to black. I have then used ordered dithering (as used by most 3D accelerators) to reduce the color depth. This dithered image is then upsampled by various filters. The first filter is a simple 2 by 2 filter, the next two are 3 by 3 filters (with different parameters). So lets look at the results:










You can clearly see the typical dithering pattern in the second normal image from the top. The 3 images below that one use a filter that tries to bring the16-bit dithered image back to an output closer to the original, 24 bit, input. As you can see, the dithering is reduced and more shades of red are created, but the dithering patterns aren't removed completely. So basically, it seems indeed true, that a higher color depth is achieved by this method. But I can't tell if the bottom 3 images are truly 22 bits, 21 bits or 19 bits.