I searched but have not found any discussion on the GR engine. That's probably because no one outside Ricoh has a sweet clue. The marketing blurb is translated directly from the Japanese. It makes a little more sense in the original (I checked) but not a heck of a lot more. The English version reads thus:

"The GR Engine was developed with the goal of reproducing images that match or surpass film camera photographs in terms of resolution and color faithfulness. Its sophisticated image processing algorithm, developed through analysis of massive sample image data, produces natural results in nearly all shooting conditions. Even diagonal lines, which are a challenge for CCD sensors, are reproduced smoothly thanks to an original image correction process in Ricoh's GR Engine. By intelligently analyzing shooting conditions and user intentions, Ricoh's new GR Engine produces exceptionally pleasing true-to-life results."

At first I was inclined to pass it off as meaningless marketing babble. In use though, its clear that the image processing pipeline in the GR Digital is significantly, perhaps fundamentally, different from any other camera, even the other Ricohs. I'd love to know whats going on in there...

Here's roughly what I've gathered so far.

Just from the brochure...

Its an algorithm, which implies the development of a different mathmatical procedure to get form the Bayer GBGR colored array of the CCD to the RGB pixels of the output image. Its nature is such that diagonal lines are rendered more-or-less correctly on the first pass without the need for an extra anti-aliasing procedure.

From experience...

Certainly with respect to the diagonal lines part, it seems to do what is written on the tin. More striking however is that apparently there is no "traditional" noise reduction at all (no blurring of textures) and no sharpening artifacts. Also, as has been noted ad nausium here and elsewhere, the remaining noise has a analog "look" to it, which, if you also de-saturate the color a little, or use the B&W setting, bears an uncanny resemblance to analog film - or if you prefer is at least sufficiently pleasing on the eye to allow the creative use of noise for artistic effect.

I tried that briefly with my new D40. Not nice at all.

Part of the answer lies in the "sharpness" control which does not produce sharpening artifacts even on the highest settings. The only effect, actually, is increased noise.

It seems the sharpening controls a notional "pixel radius", rather than the severity of a typical "gradiant highlight / edge detection" algorithm.

My guess, totally off the top of my head is someting like this:

An array of random coordinates is generated, the number of points depends on the sharpness. This net is laid over the Bayer pixel data. Each point in the net sums up color information from the Bayer pixels it is close to, the radius and relative weight depending again on the sharpness (and white balance, saturation, contrast and so forth). Finally, this random array of color grains are sampled at the final output image size, be that 8MP or whatever the camera is set to, and the JPEG compression initiated.

In B&W mode it does much the same thing but since the luminance of all the pixels can be summed together regardless of color the pixel radius can be much reduced and still give a low noise image.

If anyone has a better idea, or can suggest links to decent resources on the subject, I'd love to hear...