DOUBLE BINNABLE RGBW
Benefits
- Both 4:1 floating diffusion binning and white pixels for best in class low light performance
- Double binnable, 1/2 and 1/4 resolution modes for very high megapixel sensors
- 2x2 homogenous tiles for easy to manufacture CFA
- high quality demosaicking at all resolutions - superior to all binnable RGB CFAs
- 6 dB dynamic range advantage over RGB CFAs in binned modes and 4 dB in full resolution mode
Comparison With Existing CFAs
| CFA | Bin Mode | Demosaic Quality | Low Light SNR Advantage | Bright Light SNR Advantage | Power Consumption | Frame Rate |
|---|---|---|---|---|---|---|
| Quad Bayer | full resolution | poor | 0 dB (reference) | 0 dB (reference) | 100% (reference) | 1x (reference) |
| binned once | good | 12 dB | 6 dB | 25% | 4x | |
| RGBW-Kodak | full resolution | fair | 6 dB | 3 dB | 100% | 1x |
| binned once | good | 15 dB | 9 dB | 50% | 2x | |
| Hexadeca Bayer | full resolution | fair | 0 dB | 0 dB | 100% | 1x |
| binned once | poor | 12 dB | 6 dB | 25% | 4x | |
| binned twice | good | 18 dB | 12 dB | 6.25% | 16x | |
| Quad-IA | full resolution | good | 4 dB | 2 dB | 100% | 1x |
| binned once | poor | 18 dB | 9 dB | 25% | 4x | |
| binned twice | good | 24 dB | 15 dB | 12.5% | 8x |
TECHNOLOGY
By combining several pixels into one, binning provides the benefits of a lower resolution image sensor, such as higher SNR, speed and lower power consumption. Binning can be performed in the charge domain, voltage domain or digital domain.
When performed in the charge domain, by using shared floating diffusions, binning reduces readout noise and so provides an SNR boost beyond what is possible by voltage or digital domain binning or by reading out the whole image and downsizing it. Charge domain, or floating diffusion, binning is usually limited to 4:1 as 4 pixels in CMOS image sensors usually share a floating diffusion and its downstream read circuit.
Prior Art: Quad Bayer
Quad Bayer is the most popular binnable CFA today. However, it can only bin to 1/2 resolution which is not enough for high resolution sensors.
Demosaicking Quad Bayer also remains a challenge because of mathematical degeneracies of the CFA pattern. To learn more about these degeneracies, see section 2.2 of our paper.
Prior Art: Hexadeca Bayer
Being twice binnable, Hexadeca Bayer meets the power and speed needs of high resolution sensors. However demosaicking remains a challenge, even after one round of binning because of the resulting Quad Bayer's mathematical degeneracies.
Large 4x4 tiles of one color also makes the noise coarse grained, pushing it to the hard to remove, low frequency part of the spectrum.
Prior Art: RGBW-Kodak
The RGBW-Kodak has W pixels for improved sensitivity which gives it a 6 dB low light SNR advantage over Quad and Hexadeca Bayer in the full resolution mode. Unfortunately, the full resolution mode is not used in low light, and binning of RGBW-Kodak is limited to 2:1 compared to 4:1 floating diffusion binning for Quad and Hexadeca Bayer. This lowers RGBW-Kodak's low light SNR lead to 3 dB.
Furthermore, RGBW-Kodak can only be binned once and is not competitive in speed and power consumption with Quad and Hexadeca Bayer. Hexadeca Bayer, in particular can be binned twice to achieve 16x improvement in speed and power compared to 2x for RGBW-Kodak.
RGBW-Kodak also suffers from challenging false color issues that have plagued most, but not all, implementations.
Quad IA Binned Once
After the first round of binning, Quad-IA yields the RGBW-IA CFA that is similar to RGBW-Kodak, but without its false color problems. Broadly speaking, RGBW-Kodak's false color problems stem from all its green pixels being arranged on widely separated diagonals. RGBW-IA fixes this problem by breaking up the green diagonals and redistributing the green pixels. For a rigorous analysis of the RGBW-Kodak and RGBW-IA CFAs, see our paper.
Quad IA Binned Twice
After 2 rounds of binning, Quad IA yields a RGB mosaic and a W color plane. While the RGB mosaic so obtained is not demosaicker-friendly, the RGB mosaic-W difference image is. After interpolating R-W, G-W, B-W, the demosaicked RGB is obtained by adding W. The high SNR W color plane is also fused with the resulting RGB image to improve its SNR.