Image Algorithmics Home Page

Bayer Demosaicking

Refined by nearly two decades of academic and industrial research, the Bayer pattern has the most advanced of demosaicking algorithms. Yet significant weaknesses remain in its performance. Perhaps the most serious problems are false coloration of high frequency luminance features and artifacting on saturated colors.

The vast majority of commercial demosaickers use one dimensional estimators to generate a pair of horizontal and vertical estimates of the image and then merge the two. The merge step can be a simple selection of the horizontal or vertical estimate at each pixel, or a more sophisticated blend of the two. The artifacts of this class of algorithms - directional errors and noise streaking - are visible on most commercial raw converter outputs.

Post-processing

A host of post-processing techniques have been developed to combat demosaicking artifacts, with varying degrees of success and side-effects. Suppression of artifacts are often accompanied by smoothing out of fine features resulting in a texture-less plastic look.

The IA Solution

IA's fresh approach to demosaicking uses both a novel estimator, and a novel method of adaptively picking the estimator suitable for each image location. It is the combination of high quality estimators and benign adaptivity that yields IA's robust demosaicking performance.

Drag the yellow magnifier to zoom into your region of interest. Or download the image set and use your own viewer.

Luminance resolution

Note the reconstruction of the vertical stripes as it approaches Nyquist frequency and aliases over with very little false color artifacting.

Saturated colors

Observe the Red-Blue and Orange-Green Siemens stars being reconstructed with greatly reduced artifacting.

Image and noise texture

Note the fine features, the micro detail and the noise grain. There is no smearing and no plastic or painted look. The slightly higher noise in IA's image is due the correspondingly higher luminance bandwidth of IA's demosaicker. Being unstructured, this noise is amenable to noise reduction.

Robustness

While existing raw converters excel at some feature types, they perform poorly in others. The IA demosaicker matches or surpasses all demosaickers in nearly all feature types.

Image copyright Imaging Resource, taken with an Olympus E-M5, available in raw format here. Image selected for its exceptionally high per-pixel sharpness.