Misconceptions & Discussion on Resolution of One Shot Colour vs Monochromatic Cameras
Preface
Are you choosing a camera to use for astrophotography and can’t decide between all the options? One of the important factors to decide is choosing a one shot colour or a monochromatic camera.
OSC (One Shot Colour) cameras are typically much cheaper compared to their mono (monochromatic) equivalents, presumably due to the higher demand for OSC sensors in the general consumer camera market. Additionally, with mono cameras there is the added expense of filters: you’d need at minimum 3 filters and a filter wheel to achieve a similar RGB image. Price is one key consideration; this article, however, focuses on technical differences in the quality of data attainable.
What exactly makes a sensor a one shot colour
An OSC sensor is basically a normal mono sensor with a CFA (Colour Filter Array) superimposed in front of it. One CFA we all know and love is the Bayer matrix, which has a 2×2 pattern containing two green, one red and one blue pixel (as shown below). There are other CFAs with much more complicated designs, such as the Fujifilm X-Trans CFA often found in Fuji’s consumer cameras. However, most astro cameras with a CFA use the Bayer pattern, so we will assume our OSC camera has a Bayer matrix.
So for a given mono sensor with this CFA applied on top of it, each pixel gets filtered with a specific colour. All up, 25% of the pixels are red, 50% are green and the remaining 25% are blue.
OSC doesn’t collect as many photons, so that means worse SNR (Signal to Noise Ratio), right? No! In actuality the image signal or intensity is about the same as one taken with a mono sensor. Ignore QE (Quantum Efficiency) and responsivity for a second and assume perfect filtering by the CFA, such that a given photon can only transmit through one of the three colour filters. Then only 50% of the ‘green’ photons incident on the sensor will actually transmit, reach a pixel and generate signal, and the same for red and blue with 25% each. In reality the QE varies with wavelength, typically peaking in the green.
The CFA is also not perfect and will have transmission losses; this is shown below in the responsivity curve for a ZWO ASI224MC. It is also important to note the overlap present in some regions of the spectrum. This means a given photon incident on the sensor may transmit and reach a pixel through more than one of the three colours. A 575 nm photon, for example, could transmit through the green or the red filter. For this particular camera there is also the NIR region (800 nm+), where photons transmit approximately equally through any of the three filters.
In short, yes, a major fraction of photons incident on an OSC sensor will not generate signal compared to one without a CFA. However, it is not exactly the 25/25/50% split for RGB that is commonly quoted, and it will vary from camera to camera. A major misconception built on this fact is that “OSC cameras produce much lower SNR images”. That is only slightly true, because of the transmission losses, which are usually on the order of 5–10%; ignoring those, there is negligible difference.
For example, let’s assume we’re imaging with an Hα filter in front of our OSC camera and in front of our mono camera, and that there is no overlap in the responsivity, so Hα photons only transmit through the red pixels of the OSC camera. Only 25% of the sensor is then receiving signal compared to the mono. However, the signal of each red pixel will be the same as on the mono (ignoring minor transmission losses), and so have the same SNR. Doesn’t this mean 3/4 of our sensor will just lack signal? Yes, and this is where demosaicing (or for a Bayer filter, debayering) comes in: it interpolates values for the pixels between those that did receive signal.
Demosaicing and the real issue with OSC
The interpolation between these values is generally referred to as demosaicing an image. Many different algorithms can do this interpolation. A simple one is bilinear interpolation: a green pixel needs a blue and a red value interpolated, so the blue value can be taken as the average of the nearest 4 blue pixels, and the same for red. In our Hydrogen Alpha example, each blank pixel takes the average of the nearest 4 red pixels.
There are much more complicated means of interpolation too. One implemented in PixInsight is Variable Number of Gradients (VNG), which uses gradients in local regions to compute an estimate. VNG’s effectiveness would most likely be reduced in our Hα example, with little to no information in the blue or green pixels, possibly integrating the noise from those pixels and reducing the overall SNR, so a bilinear method may prove better. Another alternative is superpixel, where you effectively ‘delete’ the blank pixels and reduce your resolution to 1/4 of the original. Clearly we have much less spatial information with OSC, especially with an Hα filter. You can think of it as sampling the image at a lower frequency, so we should expect an OSC exposure to lack high-frequency information.
The Hα narrowband images below were taken with exactly the same equipment, settings and sensors (except for the CFA). The top left is mono, and the rest are OSC, showing the results of different debayering algorithms. All the debayered images had the red channel extracted.
These images clearly show that mono has superior resolution, and high-frequency details are lost in the OSC exposure. The interpolation quality varies, with VNG giving quite smooth results and bilinear giving harsher edges. However, the interpolated pixels are just guesses and are no true substitute for the real information a mono sensor could acquire.
Quantifying this is non-trivial, and it will vary from sensor to sensor with manufacturing tolerances and how different CFAs are made. We can quantify resolution by comparing the MTF (Modulation Transfer Function) curves of an OSC and a mono sensor. Below are some images from maxmax.com showing the MTF curves of a Canon 30D: the left shows the MTF of the mono sensor and the right with the Bayer matrix. The first row is measured with white light, the following rows with green, blue and red light. In general, the contrast drops off dramatically to near zero at the higher frequencies for the Bayer sensor. This complements what we observed earlier with actual astrophotography data.
What about dithering and drizzle?
All the discussion above has been about individual exposures. When it comes to producing an image, though, we have many tools up our sleeves, one of which is Bayer (or CFA) drizzle. If data taken with an OSC camera is dithered, Bayer drizzle can effectively reconstruct the image from the raw pixel data without any interpolation. To what extent it reproduces the lost high frequencies I am unsure, as I don’t have sufficient datasets to make an evidence-based claim. It could also be argued that drizzle integration can be performed on a mono dataset too, further improving high-frequency resolution, most likely with much less reduction in SNR. Note that in both cases drizzle only helps when undersampled.
Further considerations
Colour reproduction
As noted above, the CFA responsivity curve shows overlap between the R, G and B regions. This varies from sensor to sensor, and more importantly differs from the RGB filters used with mono cameras, which typically have square cut-off transmission curves with no overlap. Because of this, colour reproduction can vary quite a bit, and it can make a difference where light pollution is concerned.
Total integration time comparison
I’ve purposefully shied away from this topic; it is complicated. The common belief is that mono will reach a desired SNR with LRGB data faster than OSC. It is further complicated by the degrees of freedom LRGB gives you: the ratio between L and RGB, taking just RGB and synthesising L, creating a super luminance from all the LRGB masters, and so on. In theory L is much more efficient at reaching a desired SNR than OSC, as each pixel receives photons of any visible wavelength, and this L will have superior high-frequency detail. Quantifying that efficiency would need accurate QE and responsivity curves for the relevant integrals, and so will vary from sensor to sensor.
Conclusion
Hopefully it is now clear that OSC cameras are not automatically inferior to mono because fewer photons generate signal, and that there is most definitely not a 50% or 75% loss in SNR. There is a slight reduction in SNR due to transmission losses through the CFA. The major pitfall of OSC is the loss of high-frequency information in each exposure, especially with narrowband data.
Further reading
A few articles worth reading for more on some of the things mentioned here:
- Introduction to Modulation Transfer Function, Edmund Optics
- Debayer Study, maxmax.com
- Drizzle: A Method for the Linear Reconstruction of Undersampled Images, Fruchter & Hook 2001
Thanks to Frito for supplying the ASI1600MC and ASI1600MM subs.