Can RAW images be already somewhat cooked?

Tuesday March 10 2009

Sensor Insight
Fixed pattern noise correction | Noise reduction | Conclusion

The enemy of high and very high ISO is noise. In order to obtain a good image, every RAW converter includes noise reduction algorithms. The idea of noise reduction at least in homogeneous areas is simple: since noise has zero mean value (by definition), replacing each pixel by an average of the neighboring pixels (including itself) reduces the noise standard deviation. This is the most fundamental result of statistics (the law of large numbers— which also explains why opinion polls cannot be totally mistaken).

Reducing the noise standard deviation increases the signal-to-noise ratio (SNR).

How to detect noise reduction

If reducing noise is so easy, why is it not always applied? Well, there is a price to pay: averaging pixels increases SNR, but introduces some correlation between pixels. This creates a grainy aspect to the image which is often as annoying as noise itself. Moreover, if pixels are blindly averaged, fine details may simply be erased.

A few cameras apply noise reduction directly on the RAW images (for example, most Sony DSLRs).

There are at least two ways to detect noise reduction:

1. Noise reduction may be applied differently on the sensor color channels. Typically, the red and blue channels are less sensitive than the green. DxOMark Camera Sensor only reports noise measurement on the green channel because noise characteristics are the same for all channels with no noise reduction. If the SNR characteristics of the red and/or blue channels are significantly higher than the green one, then noise reduction is certainly used. This is the case for instance for the Sony A900 at a high ISO setting (ISO 6400 for the figure below). The SNR on the blue and red channels is about 6dB (2ev!) higher than the green.
Sony Alpha 900 SNR
Figure 1. Noise curves for the Sony A900 at ISO 6400 for the three color channels. The red and blue channels have been filtered and thus have a better SNR.

2. Noise autocorrelation is a mathematical tool that is a first check of the “whiteness” of a signal. If the noise is white (no local averaging), then the autocorrelation exhibits a single peak (such as in Figure 2 below) showing that each pixel is correlated only with itself and not with its neighbors. If this is not the case, then the correlation is displayed as a broader-based peak, per Figure 3.

3D view green-channel
Figure 2. 3D view of the green-channel autocorrelation for the Sony A900 at ISO 6400. The single peak shows that noise is white.
3D view red-channel
Figure 3. 3D view of the red-channel autocorrelation for the Sony A900 at ISO 1600. The peak is surrounded by a smaller dome, showing the correlation of a pixel with its neighbors