Introduction

If one is planning to use digital single lens reflex (DSLR) cameras in astronomy, it will be quite an interesting question, how the noise performance of these cameras will compete with the classical CCD. The proper answer, which camera is suited for astronomical imaging not only depends on the type of detector. It depends on the use case and the way, how standard image calibration is performed in astronomy. This leads to a surprisingly different answer when comparing cameras created and optimized for conventional photography. There also is a tendency, that DSLR cameras will supersede the classic CCD camera in the field of color devices. Manufacturers, like Sony, already announced not to develop new CCD detectors any longer.

This article will provide an overview of noise performance collected from certain dfferent digital single-lens reflex cameras (DSLR). These are compared to properties already known from the CCD technology. 

CCD, APS and the DSLR

CCD is the abbreviation for charge coupled device. In this article, the abbreviation only refers to CCD imaging sensors (there also exist other applications based on the principle of charge coupled devices, like audio signal processing). CCD detectors are used as imaging device in astronomy since decades. Their properties well understood and a dedicated image calibration pipeline has ben established as the astronomical standard for modern image processing. APS stands for active pixel sensor and is a standard technology used in many modern digital single lens reflex cameras (DSLR). To make confusion perfect, some DSLR cameras also use color CCD pixel arrays. The difference between the two sensor types, APS and CCD, is an architectural and constructively created.

The classic CCD sensor is based on MOS (metal oxide semiconductor) technology. From the collected light intensity (photo flux) a charge of few to Millions of photo electrons is created in a photo diode semiconductor device and hold by an electric field of the picture element, also called pixel. After exposure the CCD pixel images will be read out by charge transfer of the collected photo electrons through the sensor pixel by pixel. The charge is then amplified at one corner of the sensor. The obtained photo current is then converted into digital numbers to represent the photo flux. The active pixel sensor (APS) type comes with a completely different design based on CMOS technology. The APS sensor also creates electrons in a photo diode, but provides individual amplifiers with every singe pixel. The APS technology provides faster read-out, which is not only required in conventional photography. Unlike the CCD, individual amplified and addressable pixel of the APS can be read-out simultaneously and in parallel for many pixels without increasing read-out noise by increasing the pixel read-out rate. The APS technology has been remarkably improved by engineers over the last decades.

Collection of data

I personally own several DSLR cameras used for imaging and spectroscopy, in particular a Canon EOS 40D and a Canon EOS 60D in different flavors, stock cameras and astro-modified. The astro-modified cameras are used with a complete set of emission line interference filters. For comparison of the noise properties data from three Nikon models have been collected in addition: Nikon D70 and Nikon D800 owned by a friend of mine, and also a Nikon D7000 owned by an amateur astronomer.

Plausibility of the noise pattern 

In general, image noise is randomly scattered around the mean value of the illumination or black level. Depending on whether we record light or just a dark exposure with no light, there are typically two noise sources: (1) A noise pattern with a more or less symmetric Gaussian noise distribution around the mean value of the camera bias plus thermal signal (detector characteristics). (2) An additional Poisson distributed noise profile from the light source itself (natural noise from statistics of the quants of light). These two noise sources are adding together, if we record any illumination with the sensor.

A simple plausibility check will be the inspection of the noise distribution around mean value of the noise level. This is done by looking onto the intensity histogram of the non-modified, linear image intensities of the bias and dark frames. The signal should just be streched enough, so that the distribution is clearly spread. This will simply demonstrate, if the tested camera will yield a reasonable intensity distribution, or not.

Figure 1: Typical noise distribution of a long-exposure dark signal including thermal noise from the Canon EOS 60D (30 second dark exposure). Note the symmetric, Gaussian profile of the random values. Bit errors are also found from the histogram of a single exposure. Quantum conversion rate is measured at 0.6 at ISO 400. Therefore the standard deviation of noise is about half the full-width at half maximum of the noise distribution.

Evaluation of noise values

Knowing the two primary sources of noise of a dark current signal, noise values are easy to evaluate. The general procedure has been described by Berry & Burnell (2005). The technique is proven to evaluate the quantum conversion rate and noise characteristics of any astronomical CCD detector or DSLR camera.

For measuring the noise characteristics, one needs a set of exposures: One flat illuminated exposure, e.g. a "flat field" exposure taken from sky before sunset or a photo of an evenly illuminated, white screen. Two bias frames will be taken with same exposure time as the flat field exposure. Additionally, two dark exposures will reflect the thermal noise at larger exposure times, typically at exposure time of 30 seconds up to few minutes exposure.

Photon noise (light) is Poisson distributed. Hence, there exists a relation between the mean value of the recorded image intensity and the spread of its noise. Given an intensity value, its mean value and variance will theoretically be identical, if the camera is set at optimal amplification. In this case any digital value represents exactly the number of photoelectrons measured from the sensor. With any imaging detector, however, the amplification (gain) of the flux of light recorded is arbitrary and yields any ratio of the mean value and calculated variance of the noise, which is different from the value 1. This ratio of the mean value divided by the square of its variance is called the quantum conversion rate and can simply be taken off an image portion of the flat illuminated field. The noise values of the bias and dark frame can be measured by computation of their standard deviation divided by the quantum conversion rate. This yields the proper noise value expressed as number of photo electrons e^- (root mean square).

Results

Surprisingly, the modern DSLR cameras tested so far provide very low noise values of typically 3 - 5 photo electrons per pixel at ISO 400. This has been proven by independent observers and also confirmed by own measurements. The low values are true for both camera manufacturers, Canon and Nikon. Typical noise of a semi-professional or professional grade CCD detector will come with values around 8-12 electrons, with the less expensive CCD systems (typical average value from a review of current camera data sheets). Therefore, DSLRs and CCDs as the date of writing (year 2015) are comparable when it comes to noise, with the trend of noise performance pointing to lower values especially for the APS CMOS imagers.

The average noise signal level of the dark signal will be internally regulated by the camera electronics to any (negative, zero or positive) value. Usually, this is done for example by adding an arbitrary offset voltage (analog) or computed after digitization. Some imaging sensors provide unexposed image areas, which can be used to compute a bias value from "black" pixel values. Certain CCDs used in the past also created digital signal data format (files) containing negative values. However, DSLRs like the Canon or Nikon models don't provide negative values with their raw file formats.

The tested Nikon D70, D7000 or D800 cameras come with certain drawbacks in astrophotography. Nikon creates an artificially low drawn bias close to or equal to zero level. This means, that the noise distribution will not form a Gaussian distribution. Instead, negative values around the zero mean value are just clipped (again, the analog-digital converter only provides positive values from the raw frames). It is very likely, although not certain, that the Nikon model clips negative values, but will not process measured intensities by using noise suppression, if switched off. Evidence is given by observation of a spike in the histogram at zero, which supports the assumption of clipped negative numbers set to zero. This kind of clipping does not yield true physical noise values with standard computations, because the signal is showing a half-shaped Gaussian profile. Therefore, the true bias level and variance will not be directly accessible to standard photometric analysis. Instead, any computed standard deviation of the clipped noise values will be under-estimated and the mean value will be over-estimated because of missing (clipped) intensity values of the noise distribution. May such computations be suited to claim a low noise signal for the Nikon, it is not a measure of the true detector noise. For serious astronomical image processing this yields a wrong absolute photometric value of the bias and a wrong photometric error, as well.

 

Figure 2: Typical noise distribution of the Nikon D7000 from a short-exposure bias frame. Clipping of negative noise values at zero level yields a half-shaped Gaussian noise profile with no negative values below zero, which leads to false estimation of the true noise properties. Like other imaging devices, the Nikon models also show typical bit errors from the analog-digital converter unit.

Decreasing the average bias level by internal regulation of the electronics will mean a gain in overall dynamic range and create stable image properties. This is standard to both, the CCD and DSLR cameras. From the evaluation of both Canon models, a bias level has been found at around 1024 for the Canon EOS 40D and a value of 2048 for the Canon ESO 60D. The largest intensity value, that can be stored in the raw frames, is 16383 with a 14 bit analog-digital converter unit. As the standard deviation of the detector noise is shown to be very small, such large bias values create a huge buffer not to clip values. Larger than that, what would be required to balance the bias signal to avoid clipping. A bias of 3 times the spread of its noise (standard deviation) already would mean a probability of 99 percent not to accidentally clip values. The higher bias levels of both Canon models are more than just safe, but cause a slight loss of the possible dynamic range available. On the other hand, the distribution of the noise of both Canons is more suited for serious astronomy, because they do not clip or modify any measured intensity values. From the simple plausibility check of dark frames obtained, the Canon models tested will better support scientific photometry, than the Nikons.

It shall be noted, that intensity values from the DSLR raw frames shown in the above histograms are corrected to reflect a 16 bit detector. Thus, intensity values from the raw image files are multiplied by 4 before the signal processing takes place. This correction in my software ArgusPro will be done for historical reasons and to normalize typical 12, 14 bit DSLR frames and also support 16 bit frames taken from CCD detectors. Therefore, the measured bias and noise levels will appear four times larger from the screenshots, than the real values taken out of the raw images are.

Other properties

One shall keep in mind, however, that requirements for astronomical imaging are different from conventional photography. Low noise characteristics are beneficial for low light level astrophotography, as random noise from the detector will have a serious impact to photometry and limiting magnitudes of the faint objects recorded. It will always be better to have plain noisy values from the detector to be able to chose whatever algorithm will perform best afterwards for any given use case, like photometry, post-processing or image deconvolution in the astronomy domain.

Active pixel sensors have a few properties in common with the classical CCD detector. These are a number of hot pixel, probably also dead sensor pixels. These can be found as brighter (darker) pixels in bias and dark current frames. They can be easily removed by subtraction of the measured (average) dark signal from an astronomical exposure. However, the astrophotographer will find these hot pixels moving with time, i.e. by using a DSLR. Therefore, dark frames shall be recorded with every new observation to control temporal effects of the dark signal and bias level. However, this temporal movement of hot pixels has also been found with certain CCD cameras, which the author used in the past.

Bit errors from the analog-digital converter (ADC) unit of the camera can be found with both, the CCD and APS designs. The bit errors are easily found as periodical gaps in the intensity histogram. They occur, because the number of digital bits set to 1 (one) or 0 (zero) will mean different current drawn from the power supply of the ADC, which may cause voltage drops and interfere with the amplifier section of the detector. In other words, there will occur voltage coupling between the number of 0/1 bits and a floating pixel bias voltage level measured. The effect can be shown to disappear by averaging several frames and will not have a serious impact to photometry.

There are also many other effects found with APS detectors, that are already described in the literature to CCD astronomy, like detection of high energy particles with the optical sensor. These sometimes are found as short traces or bright spikes in the image, sometimes found in two or more consecutive frames.

Conclusion

This article is meant as a lecture how to measure and understand the noise properties of a DSLR camera used for astronomy. I personally didn't find obvious hints, that any of the cameras tested so far will do some magic to noise suppression, if internal suppression of noise is switched off. It is just noise raw signal written to digital images, if the camera is carefully set. However, the advice to carefully switch off any internal noise suppression or automatic calibration or color adaptation of the camera shall be given to DSLR users. How to do so will be described in the respective owners manual of the individual camera model.

One shall keep in mind, that requirements for astronomical imaging are different from conventional photography. Thus, a camera optimized for conventional photography must not necessarily be useful to serious astronomy. Low noise characteristics are beneficial for low light level astrophotography, as random noise from the detector will have a serious impact to photometry and limiting magnitudes of the faint objects recorded. It will always be better to have plain noisy values from the detector to be able to chose whatever algorithm will perform best afterwards for any given use case, like photometry, post-processing or image deconvolution in the astronomy domain.

APS cameras, like the Canon or Nikon models, provide imaging characteristics very similar to the classical CCD detector. In comparison to the CCD, surprisingly low noise values are found with the DSLR, but the properties of the noise profile recorded may differ between manufacturers. Therefore, this article cannot be meant as a competitive test, which manufacturer creates the better cameras with regards to the true noise properties. It will be the special requirements and demands of astronomical image processing, that will yield the proper answer.

Care shall be taken when doing photometry of the very faint objects by using cameras introducing artificial clipping of intensities. This will have a negative impact to serious photometry, but might be tolerable for standard astronomical imaging. It should be taken into account, however, that even the standard calibration procedure of subtracting a dark frame from the image might introduce artifacts at the low noise level close to the limiting magnitude due to clip of intensity values.

 

Literature

Berry, R. and Burnell, J.: The Handbook of Astronomical Image Processing. Willmann-Bell, 2005.

Buil, C.:  Comparaison des Canon 40D, 50D, 5D ET 5D Mark II. [ Link ]

Centen, P.: CCD On-Chip Amplifiers: Noise Performance versus MOS Transistor Dimensions. IEEE Transactions on Electron Devices, 38, No. 5, 1991 .

Fish, A., Yadid-Pecht, O.: Active Pixel Sensor (APS) Design - From Pixels to Systems. Springer, 2004. 

McLean, Ian S.: Electronic Imaging in Astronomy: Detectors and Instrumentation. Springer, 2008.