Article on Yahoo about new Raw from Adobe

Whats the big deal of a universal format? All the big players' raw format is supported by the major softwares out there. Photoshop's RAW plugin reads files from numerous cameras.

It makes it EASIER for Adobe to incorporate RAW into Photoshop. Also, people purchasing newly released cameras won't have to wait for Adobe or their favorite editing software authors to write a converter for the latest and greatest RAW format.

Adobe won't have to write file converters for all the different cameras. There are getting to be quite a few RAW formats.. Canon alone has several.. Adobe has to pay software writers to keep up with this. As a matter of fact, they had to write yet another routine for the 20D's RAW format because it isn't compatible with the 10D's RAW format..

The idea is that in the future camera manufacturers will make cameras that create DNG files instead of proprietary RAW format files. The DNG format is a kind of standard RAW format. As you probably know are RAW files much smaller than TIFF files because in a RAW file, only one colour per pixel is stored (just like the sensor records the data, because of the Bayer pattern).

leony wrote me in a private message (Why are you sending me this in a private message instead of posting it in the forum?):

It may be a surprise to you, but you're wrong. There are NOT three photo cells for each pixel in a sensor with a Bayer pattern, as most cameras have. Of each group of four pixels, two record only green, one records only blue and one only red. The RAW conversion software (in the camera or on the computer) interpolates the missing colours for each pixel to make an RGB image.

So a 6 MP EOS 10D for example has only 1.5 million red, 1.5 million blue and 3 million green pixels. There are NOT really 6 x 3 = 18 million photo sites on the sensor, as you suggest.

Search the web for more info on Bayer sensors.

That's true, RAW files are compressed with a lossless, ZIP-like algorithm. But they also contain only one color per pixel, see above.

leony, here are some articles about how Bayer sensors work:

Anatomy of a Digital Camera: Image Sensors
How Digital Cameras Work - Capturing Color: Bayer Filter
Image Sensors from photozone.de

Specifically, your line "each three photo cells on the sensor comprise one pixel that is saved to the file. RAW does not store info about every single photo cell" is not correct. Each pixel indeed has only one photo cell and in the RAW file, the info for each photo cell is stored.

The sensors of most digital cameras record 12 bits per pixel, so the uncompressed RAW data of the 10D is 3072 x 2048 x 12 bits / 8 bits per byte = 9,437,184 bytes. By compressing it with a ZIP-like algorithm, the RAW file is usually between 4 and 6 MB in size.

An uncompressed 8-bit TIFF file contains 3072 x 2048 x 8 bits per channel x 3 channels / 8 bits per byte = 18,874,368 bytes, twice as much as the uncompressed RAW file. That makes the RAW file already twice as small, even without compression.

Jesepr:

you are correct about the way info is read out from the sensor and the way sensors work - but that, however does not mean ALL of the info is saved to RAW unprocessed. Here's my logic, and I might be wrong - in the interests of science...

if, as you correctly say, each pixel that is on the CMOS is comprised from 4 photocells, and the "visible" image pixels are then interpolated from those 4 cells for each pixel photoshop considers an RGB value, to acheive a file of 2048x3072=6291456 pixels in visible image.

lets work backwards... my 10D is 6.3 MP - that is there is 6,300,000 photo cells on it. since each photo cell is a 12-bit value it takes 1.5 bytes to store. assuming there is no overhead (for easy math) you end up with a 9 MB file which after compression becomes about half the size - in theory.

so, here you have this 9MB file. and you think it has all those photo cells info, except it doesn't. what it has is info DERIVED from interpolating photo cell information. each "visible" pixel becomes from weighted averaging of a set of 9 adjacent pixels for RGB value.

Now, it is possible that canon lets this interpolation be done by software when you convert RAW, but it is just as easy to implement this basic interpolation at hardware level especially that on CMOS each pixel can be accessed individually. the overhead at hardware level involved is minimal.

now, since these photocells are a set of 4 to comprise 1 visible color pixel in RGB model, the effective "resolution" of 6.3 MP sensor in "visible RGB" pixels is 6.3/4=1,575,000 pixels. since that image is too small for any practical use, canon interpolates it to fill the area of an image pixel by pixel in 9 pixel square (3x3) increments. this is called "bycubic inteprolation" that photoshop uses to upsample files...

bayer matrix only describes how photocells are located on the actual chip. it does not describe how the chip is read and if the info is then "interpolated" before saving to a file.

in all reality - bycubic interpolation of the sensor's photocells is a necessary step to have an RGB image - be it done in camera before RAW is saved to a file or after during RAW conversion. only canon knows, but as you can see file sizes have nothing to do with this.

it is also not surprising that RAW is smaller than TIFF - raw holds info for 6291456 cells each cell at 12 bits = 9437184 bytes.

TIFF with 6291456 "visible pixels" has to hold 24 bits (R+G+B: 8*3) or 48 bits (16*3) with file sizes of 18.8 Mb and 100 Mb files respectively.

note that your RAW file that has photocells info is only 1.5 million "visible pixels" - all the other pixels are interpolated to create a 6 million pixel RGB file.

last, in deffence of my theory, a B&W (grayscale) file saved as TIFF uncompressed provides 12 and 6 Mb files in 16-bit and 8-bit respectively. see, after compression you still get a TIFF of those dimensions comparable to RAW - once you note that photo cells store "color blind" values - just like B&W. if there would be a way to save a file as 12-bit in B&W, file after compression would be exactly the size of RAW.

RAW doesn't hold any more information than TIFF does - it just holds the same information in a different way, a way not usable until conversion to RGB model where color is determined by 3 variables, not 4.

I agree that a very similar result can be had using either method and didn't say (or intend to say) that you are losing data in a RAW to TIFF conversion. What I meant was that some of the functions available in Camera RAW are more powerful than their 'equivalent' in PS. The major one being the white balance setting... a few others are slightly better or equivalent, while others are worse. I don't know the inner workings of the two systems, this is simply information I read from a published book on using Camera Raw... and I choose to believe it.

No, I didn't say each pixel is 4 photocells and that isn't correct either. Each pixel is not 4 photocells, but 1 photocell. There are 3072 x 2048 = 6,291,456 photocells = pixels in the image. The trick is that each of those pixels has only 1 color instead of 3. The missing colors are calculated from the neighbouring pixels.

No. The RAW data is what comes from the photocells = pixels on the sensor directly. The data in the RAW file is not yet interpolated. The RAW conversion software does the interpolating.

To make an RGB image from the RAW data, the conversion software looks at each pixel. For a pixel that contains only an R value, it calculates the G and B value for that pixel by interpolating the values of the surrounding pixels (and for each G pixel, R and B are interpolated etc). The most simple algorithm for doing this is calculating a weighted average, but there are also far more sophisticated (and complicated) algorithms available, which are supposed to give better results.

When you set the camera to RAW mode, the RAW, uninterpolated data is saved in the CRW file without interpolation. When you set the camera to JPEG mode, the camera / sensor does the interpolation itself.

There are not 4 photocells per pixel, see above.

Right, that's what I said in my post above, except that a 16 bit per channel TIFF file is not 100 MB but twice that of the 8 bpp file: 37.7 MB.

No, it does not have 1.5 million visible pixels, it has 6 million pixels but each with only one color and the other colors for each pixel have to be derived (= interpolated) from the surrounding pixels.

If the RAW file contains info on only 1.5 million "visible pixels", as you say, then you're saying that the 10D is in reality an 1.5 MP camera and the 6 MP image is created by upsampling (bicubic interpolation). That's not true. The 10D produces an image with 1.5 million red, 3 million green and 1.5 million blue pixels. That's not the same as 1.5 million RGB pixels.

Correct.

i understand how this all works now, and still am amazed that a necessary (and really an unchangeable) step is not done in camera.

As I wrote above, when you set the camera to JPEG mode, the camera does the interpolation.

The interpolation step is not really "unchangeable" - as I also wrote above already, there are several algorithms available to do it. The most simple way would be to just take the mean value of the missing colours. There are also more sophisticated algorithms available, which are supposed to give better results than simply calculating the average. Here's a page that compares different RAW interpolation algorithms: A Study of Spatial Color Interpolation Algorithms for Single-Detector Digital Cameras.