Image Algorithmics Home Page

The Bayer and other commercial CFA patterns have been empirically optimized. Recent research has allowed IA to mathematically analyze CFAs and methodically optimize them. This optimization is carried out with the 3 primary objectives of spatial spectrum shaping, spatial frequency assignment of luminance, chrominance signals to minimize crosstalk between them and minimization of noise amplification during demosaicking, as well as with the secondary objective of maximizing light transmissivity through the color filters. The CFA pattern shown below is the culmination of these efforts.

This pattern differs from the Bayer pattern not only in the filter colors but also in the geometry of pixels. Instead of square pixels, the proposed pattern has rectangular pixels of the same area but aspect ratio of 1.41:1. This is the same shape as an A4 sheet of paper. The output image has the same resolution in all directions and this resolution is equal to the horizontal Nyquist limit of the sensor and greater than the OLPF limited resolution captured by leading Bayer cameras. The finer vertical pixel pitch of the sensor is not used to capture extra vertical detail, but is instead used to capture color information. The output image is in square pixels so that the rectangular photosite shape is internal to the camera and not exposed to the user.

In comparison to the Bayer CFA with the same pixel count, similar luminance and chrominance resolution, and demosaicked with the popular AHD algorithm, the proposed pattern has the following advantages:

  1. 7.6dB PSNR improvement on the Kodak image set
  2. Even greater, 10dB PSNR, improvement on realistic images
  3. Greatly reduced artifacts
  4. One quarter as much increase in MSE due to noise. This is because of:
    • lighter filter colors that let in more light
    • no directional sensing algorithms to get confused by noise
    • numerically stable demosaicking, unlike CMY patterns
  5. Uniform light transmittance of all 4 filter colors - hence resistant to sensor saturation
  6. Low complexity demosaicking using separable filters, can be used in phone camera video
  7. Simple noise characterization resulting in more effective post demosaicking noise reductinon
  8. Chromatic Aberration correction can be performed after demosaicking with an order of magnitude less increase in MSE than Bayer

Top row: Original image. Second row: a small patch of the original image and as it’s captured by the proposed CFA, Bayer AHD, Bayer POCS respectively. Third row: respective errors amplified by a factor of 10. Fourth row: respective errors amplified by a factor of 4 after addition of noise. Bottom row: respective errors amplified by a factor of 6 after Chromatic Aberration correction