Video and Image Bayesian Demosaicing With A Two Color Image Prior Microsoft Research, Redmond, WA Abstract The demosaicing process converts single-CCD color representations of one color channel per pixel into full per-pixel RGB. We introduce a Bayesian technique for demosaicing Bayer color filter array patterns that is based on a statistically-obtained two color per-pixel image prior. By modeling all local color behavior as a linear combination of two fully specified RGB triples, we avoid color fringing artifacts while preserving sharp edges. Our grid-less, floating-point architecture can process both single images and multiple images from video within the same framework, with multiple images providing denser color samples and therefore better color reproduction with reduced aliasing. An initial clustering is performed to determine the underlying local two color model surrounding each pixel. Using a product of Gaussians statistical model, the underlying linear blending ratio of the two representative colors at each pixel is estimated, while simultaneously providing noise reduction. Finally, we show that by sampling the image model at a finer resolution than the source images during reconstruction, our continuous demosaicing technique can super-resolve in a single step. People Eric P. Bennett, Matthew Uyttendaele, C. Lawrence Zitnick, Richard Szeliski, and Sing Bing Kang Publications Bennett E., Uyttendaele, M., Zitnick, L., Szeliski, R., and Kang, S.B. Video and Image Bayesian Demosaicing with a Two Color Image Prior. In Seventh European Conference on Computer Vision (ECCV 2002), volume~1, pages 508-521, Graz, May 2006. Springer-Verlag. ECCV Pre-Print PDF (1.1 MB) Contact If you are interested in further details, please contact Eric Bennett. Supplemental Material

Because the ultimate metric of a demosaicing method is how good it looks to a human viewer, we present large format versions of the results shown in the paper.

To use the supplemental materials, simply roll your mouse over each image to see our result appear. This way, you can perform an A/B test with aligned images, which is easier to judge than the side-by-side images in the paper. The better of the two results is always the one that appears with the mouse over the image.

Sections:

Demosaicing Example #1 (Enlarged 'Q')

Demosaicing Example #2 (Enlarged Crayon Tips)

Super-Resolution Examples

NOTE: Images are uncompressed as to not introduce noise other than demosaicing artifacts.

1) Enlarged 'Q':

Original Image
Original Resolution: 1752x1168

Here is the enlarged version of the Q at the top of the image. The original image is shown with each demosaicer as a roll-over. Note the severe aliasing and color fringing the in bilinear approach. The aliasing in the High Quality Linear Interpolation (HQLI) image is better, but our method further removes aliasing in the 1 and 3 input image cases.

Original vs. Bilinear Interpolation

Original vs. High Quality Linear Interpolation

Original vs. Two Color Demosaicing with 1 Input Image

Original vs. Two Color Demosaicing with 3 Input Images

Here is our "bootstrap" algorithm, HQLI, shown versus both runs of our algorithm on roll-over. The difference is especially evident in the 3 input image case, with its cleaner edges with less discoloration.

High Quality Linear Demosaicing vs. Two Color Demosaicing with 1 Input Image

High Quality Linear Demosaicing vs. Two Color Demosaicing with 3 Input Images

Finally, here is a comparison of our 1 input image technique, versus our 3 input image technique on roll-over.

Two Color Demosaicing with 1 Input Image vs. Two Color Demosaicing with 3 Input Images

2) Enlarged Crayon Tips:

Here is the enlarged version of the crayons in the middle of the image. The original image is shown with each demosaicer as a roll-over. The bilinear approach over-smooths the image as well as introduces color artifacts. HQLI is sharper, but adds dark artifacts on the orange and yellow crayons.. Our techniques are both sharp and exhibit accurate coloration.

Original vs. Bilinear Interpolation

Original vs. High Quality Linear Interpolation

Original vs. Two Color Demosaicing with 1 Input Image

Original vs. Two Color Demosaicing with 3 Input Images

Here is our "bootstrap" algorithm, HQLI, shown versus both runs of our algorithm on roll-over. Again notice how, even though our algorithm is being seeded with edge artifacts on the yellow and orange crayons, we are robust and still able to properly reconstruct the colors.

High Quality Linear Demosaicing vs. Two Color Demosaicing with 1 Input Image

High Quality Linear Demosaicing vs. Two Color Demosaicing with 3 Input Images

Finally, here is a comparison of our 1 input image technique, versus our 3 input image technique on roll-over.

Two Color Demosaicing with 1 Input Image vs. Two Color Demosaicing with 3 Input Images

3) Super-Resolution Examples:

Original Image
Original Resolution: 1752x1168
Upsampled To: 3504x2336

Close-ups of our super-resolution results. For each image pair, the left is a bicubically upsampled single input demosaiced image from our demosaicer. The right image is reconstructed using our three image demosaicing technique on a denser sampling grid. Notice that our results are less blocky.

Bicubically Upsampled	Super-Resolution	Roll-Over