Sensor Analysis Primer White Level
Prepared 2009-03-01, last revised 2009-03-02 by Bill Claff

Introduction

At high signal levels the Photon Transfer Curve drops off sharply because of noise being clipped on the right-hand-side.

The point at which the drop-off occurs is called the White Level.

Determining White Level

The observed noise drops of beyond the White Level because the noise distribution is truncated on the right.

Another effect of noise distribution truncation is that the signal is also underestimated.

Therefore, beyond the White Level; the sensor no longer has a linear response.

With sufficient Photon Transfer Curve data, the White Level corresponds to the signal with the maximum amount of noise.

White Level

Based on the D300 14-bit Photon Curve Data I measured a White Level of 26,679 electrons.

We can summarize the 14-bit ADC clipping point and White Level (WL) values for the D300 various ISOs in units of electrons as follows:

 

ADC clips

WL

 

ADC clips

WL

 

log2(electrons)

 

electrons

Lo

14.81

14.70

 

28,744

26,679

200

14.48

14.37

 

22,814

21,175

400

13.48

13.37

 

11,407

10,588

800

12.48

12.37

 

5,704

5,294

1600

11.48

11.37

 

2,852

2,647

3200

11.31

11.20

 

2,541

2,358

6400

11.31

11.20

 

2,541

2,358

 

Note that White Level stops dropping at high ISOs because these ISOs are accomplished through digital scaling.

Finally, in terms of 14-bits ADUs, the ADU clipping point is 16,384 and the White Level is 15,207.

Discussion

The ADC clipping point based on gain is 28,744.

The White Level is further below the ADC clipping point than we would expect based on Photon Noise alone.

That is because even a strong signal has more than Photon Noise.

The chief other noise source is Photo Response Non-Uniformity (PRNU).

PRNU is why we used pairs of images to determine gain.

 

Note that the White Level is .11 EV (1/9 EV) below the ADC clipping point.

 

Although the sensor is not linear beyond the White Level, with proper calibration; corrected values can be determined.

In fact, a properly calibrated sophisticated decoder can recover highlight values beyond the ADC clipping point.

Given this fact, its unclear whether White Level or Full Well Capacity should be used as the upper bound in Dynamic Range calculations.

As a practical matter, because determining White Level requires a great deal of data, I use Full Well Capacity in my Dynamic Range calculations.