> Date: Fri, 16 Jun 2006 22:40:15 -0400 > From: John Francis <[EMAIL PROTECTED]> > Message-ID: <[EMAIL PROTECTED]> > > On Fri, Jun 16, 2006 at 09:47:29PM -0400, Igor Roshchin wrote: > > > the difference in number is negligible (in the scenario described > > here it would be ~ 5000 pixels for a ~2000x3000 matrix which is ~0.1%, > > but in reality it could be a bit more (no pun intended), depending on > > the implementation) > > Not only is this over-simplified; it's wrong. Most de-mosaicing > algorithms put the pixel centres coincident with the sensor sites, > rather than in between. They don't replace an actual sensor value; > they interpolate missing components, often using rather more than > just the four neighbouring pixels. Because they use values from > one or more rows or columns of neighbouring pixels is one reason > why the 'full' image size (3008 x 2008 in this case) is smaller > than the actual sensor dimensions (3040 x 2024).
Indeed, what I wrote is a simplification. I did not intend to present all possible types of demosaicing schemes. I wanted to demonstrate the concept of what happens in the RAW conversion to the extent of why RAW format has advantages over TIFF/JPEG. Sorry, I meant to write explicitly that it was a simplified picture, but now realized that forgot to do so. Yes, I know that the interpolation from more than four neighboring pixels can be used, and acknowledged this in the part quoted above. The Bayer filter mosaic (used in my example) is not the only one possible. And there are many other details that I didn't mention. More over, I intended but forgot to mention in the end that there are sensor designs where different color sensors are all stacked in one photosite. But that is very diffirent type of sensors. > > > It is this demosaicing process when the white balance, sharpness, gamma > > correction, ISO values, etc are being used in the conversion. > > Again, incorrect. The ISO value is generally used to control the > gain of an analog amplifier during the actual data readout from > the sensor before the conversion to digital values. As such, the > sensor values in the RAW file have already had ISO scaling. In fact > in some cameras (although not in the Pentax models) the white balance > is also applied in the analog domain. Furthermore, sharpening at > this stage is only done in certain converters; while Adobe Camera > Raw can apply sharpening (and image re-sizing) during RAW conversion, > many converters do no such processing, leaving that for editing. Yes, some of those values are (can be, depending on the design) applied even before RAW conversion. It was a somewhat frivolous for add ISO to this list. My point was that those values (with a limit in case of ISO) can be changed in the RAW conversion, but they are mostly lost after the conversion. > > > Since the process of demosaicing is fully mathematically reversible, > > Not necessarily. While a simple averaging scheme is (close to) > reversible, subject only to there being not quite enough data values > to uniquely determine the outermost pixel values, there are several > other demosaicing algorithms that are not reversible. Oops, it was a typo. "not" got lost. You could've guessed from the entire sentence. "Since the process of demosaicing is NOT fully mathematically reversible, the loss of the initial information happens at this point." Thank you for noticing it. > > > I think this rather simple documents explains some of these issues > > in more detail: > > http://www.adobe.com/digitalimag/pdfs/understanding_digitalrawcapture.pdf > > > > I hope I didn't cause too much confusion, and my explanation was > > helpful to somebody. > > So, I stand corrected where it is due (thanks, John!), and still hope that even despite possible confusion my message was helpful to somebody. Igor -- PDML Pentax-Discuss Mail List [email protected] http://pdml.net/mailman/listinfo/pdml_pdml.net
