An in-depth case study of the use of DxOMark data

Tuesday June 09 2009

Camera Article
Introduction | Performance overview | RAW noise analysis | Color blindness & sensor quality | Dynamic range and noise source | Conclusion

The primaries of the sRGB color space do not correspond with the primaries of the sensor (see Color measurements), and thus the RAW values for the three color channels do not match the RGB values of the output color space, (i.e., sRGB). Each manufacturer has to apply color processing that tries to match the sensor response to output color space. The simplest solution is a color matrix and a tonal curve.
This color processing applies some gain to the different channels and mixes them to obtain a satisfying final color rendering. This processing is highly dependent on the sensor spectral response (how the sensor behaves when exposed to a particular wavelength). On the one hand, if the spectral responses are very wide, many subtle color nuances might not be distinguishable; on the other hand, if the spectral responses are not wide enough, the sensor will not be sensitive, yielding noisy images.
It is also well known that evaluating only SNR is an insufficient basis for comparing RAW sensor behavior, which is why DxOMark measures color sensitivity to further predict noise after color processing.

Measurement Results

Below are the two camera’s sensor results in print mode:

Color sensitivity (mode print)

This comparison is a bit surprising with respect to the previous SNR 18% results. Why such a difference? Color sensitivity is impacted by noise curves and spectral responses. If SNR curves are close, most of the divergence observed must be due to a difference in spectral sensitivities, which implies very different color processing for each sensor.

Please refer to the color response tabs for each sensor; below are the results for a D50 illuminant:

Color response, Canon EOS 500D
Color response, Nikon D5000

As explained above, color processing is always applied to match sensor response to the sRGB response. Color measurements describes how color processing mixes raw color to obtain real sRGB color. For example, a red channel is expected to contain mostly red, a bit of green, and almost no blue. From the EOS 500D graph above, however, it is pretty clear that the red channel does not follow this principle.

Conclusion

Despite very similar SNR curves, color depth shows a wider difference between the two cameras than expected due to the different spectral responses of their sensors (see “further information” below). These differences imply two different types of color processing—i.e., the Nikon’s color matrix uses a low coefficient, but the lack of purity of Canon’s red channel requires that the color matrix apply a higher coefficient to compensate.

Color depth measurements increase the gap between the Nikon and the Canon for low ISO. Although the Canon EOS 500D’s SNR curves are better than the Nikon D5000’s for high ISO, thus closing the gap to some degree, the net result is that the Nikon D5000’s score is still better.

Further information

1. Relative sensitivity

All the standard RGB color spaces (such as sRGB, Adobe RGB, Prophoto) assume that a neutral reflectance object (i.e., an object that reflects every wavelength equally) obtains equal values on the three channels. However, the RAW values on the sensor depend on the spectrum of the light illuminating the neutral object, and on the spectral responses of the sensor. The responses of the three channels are usually different. The relative sensitivity of both the red and blue channels is the ratio between the value of the red (or blue) channel and the green channel (see Color measurements).

The graphs below show the relative sensitivity of the Canon and Nikon sensors for a D50 illuminant:

Canon EOS 500D

Nikon D5000

Color_500D_RelativeSensitivity

Color_D5000_RelativeSensitivity

The two results are similar: for both cameras, the white balance applied to compensate for the red and blue channels’ lack of sensitivity in comparison with the green channel is very close.

2. Spectral sensitivities

Channel decomposition is a low-level description of the sensor spectral response. The easiest way to explain the following results is to measure the spectral sensitivities of the two sensors:

Canon EOS 500D

Nikon D5000

Color_500D_RelativeSensitivity

Color_D5000_RelativeSensitivity

Of interest here is the spectral sensitivity of the Canon EOS 500D red channel, which is not as selective as the Nikon D5000’s.

As proof, look at the sensitivity level of the blue and red channel at 550 nm. 550 nm corresponds to a pure green wavelength, meaning that the blue and red channels should have very limited sensitivity to this wavelength.

While the Nikon red channel is at only 3%, the Canon is still at a very sensible 26%. Same for the blue channel: Nikon is at only 2%, while the Canon is at 12%. Canon clearly tried to push up the global sensitivity of the sensor by applying a less selective filter, but its channels are not selective enough and the need for stronger color processing implies greater color noise (see “Further Information” below).

This lack of selectivity introduces a high overlap between the green and red channels, impacting the color matrix as shown below:

3. Raw channel decomposition

Per the graphs following, the green and blue channels of the two sensors behave similarly:

Canon EOS 500D

Nikon D5000

Green

500D_Green_primary

D5000_Green_primary

Blue

500D_BluePrimary

D5000_Blue primary

The two sensors’ behavior changes on the red channel:

Canon EOS 500D

Nikon D5000

Red

500D_Red_primary

D5000_Red_primary

The red channel of the Nikon D5000 shows a typical behavior, but the red channel of the Canon 500D presents a very high mix of red and green. As its sensor is not able to distinguish the little nuances between green and red, the color rendering algorithm of the Canon RAW converter software has to compensate for this lack of color accuracy by applying high gain during color processing.

4. Color matrix

To simplify, the color matrix is applied to match the sensor color response to the color response of the human eye. But neither of these color responses matches sRGB primaries.

So the color matrix gives the coefficients of the linear combination of the sensor channels, which are then applied to compensate for the sensor channels’ lack of purity as compared to the sRGB primaries. The greater the differential coefficient, the more the noise will be increased by color processing, as shown by the results for a D50 illuminant below:

Canon EOS 500D

Nikon D5000

ColorMatrix_Canon

Color_Nikon

Here we see that the Canon EOS 500D has very different color behavior than the Nikon D5000. The red channel is not selective enough, so the color matrix has to apply a stronger color processing to mix the different channels to obtain a satisfactory match with the sRGB. In other words, the Canon EOS 500D is a bit more color blind in comparison with the Nikon D5000. To compensate, a greater degree of color processing is applied, which increases noise.

This will impact the final RGB picture even if noise processing reduces the effect. Generally speaking, even with a good RAW converter, very low frequency colored noise remains on the final image.