Shrinking pixel sizes have reduced both the radiant energy per pixel and full well capacities. While RGBW CFAs improve low light performance by filtering out less light, this does not help bright light SNRs which are limited by well capacities.
A solution to this problem is to resist clipping of bright areas thereby increasing the dynamic range of the system. This increased dynamic range can be used to faithfully capture brighter objects, or can be traded for better SNR by over-exposing the sensor.
An old idea for resisting highlight clipping is to fall back on RGB only demosaicking if the brighter W pixels clip. While sound in principle, this places even greater demands on the demosaicker than non-clipped RGBW.
IA offers a solution that leverages its advanced demosaicking technologies to perform higher quality RGB only reconstruction than was previously possible. This is combined this with processing tailored to the unique image characteristics of bright areas to further improve highlight quality.
IA's processing trades off overexposure performance with artifacting and can tolerate 6dB of overexposure on typical images with little artifacting, provided blooming is controlled by the sensor. This enables the RGBW CFA to equal the noise performance of a Bayer sensor with twice the pixel area.
Drag the yellow magnifier to zoom into your region of interest
Observe the uniform texture and the non-processed look of the image. Hightlight recovery is seamless.
Note the color uniformity between the clipped and non-clipped image areas. This is a possible since the IA demosaicker does not rely on W for color or assume R+G+B = W.
Clipping artifacts are most visible on high contrast edges since the aliased energy is greatest there. Note the improved edges of the green leaves against the clipped white background; they are significantly blurred in the non-recovered image.
Download more images here. All images are from the IMAX image set and mosaicked with the Truesense RGBW CFA with noise modeled on the KAI16070 sensor. Test conditions for generating the image set is outlined in the included Readme.txt file. Comparable performance has been demonstrated with IA and Sony RGBW CFAs.