Spectral Cognisance Part 1 – Bad Red or True Blue?

Here is a real-life event inspiring a blog post: One of my team’s previous sensor projects is being implemented in our internal camera development. We had just received a new lot of sensors which appeared to have low Responsivity. We soon realized that an IR-cut filter was left in the test system from some time back when a colour camera was under test. Not surprisingly, with the proper test configuration, the new devices showed the Responsivity we had expected.

This episode highlights the importance of light sources, filters, and spectral considerations in general in optical systems and inspired me to start this series of posts.

If you know a bit about black body radiation (as emitted by our halogen light source) you realize that an IR cut will have a major impact on the actual number (and spectrum) of photons reaching the sensor.

Consider the following plot – it shows the photon emission spectrum of the light source at 3500K (black graph), the same spectrum after the IR-cut filter (red graph) and a typical QE curve of a CMOS image sensor (blue graph).

Photon Emission & Absorption Spectra - Black Body & IR-cut

Figure 1: Black Body Emission with/without IR cut & Sensor QE

We measure with a light power meter that also uses a photon detector – it will have a QE curve that is different from the one shown for a CMOS Image Sensor. The meter assumes all signal comes from a single (programmable) wavelength and applies a (calibrated) gain factor accordingly. Its output is a light power value in µW/cm2, or – taking into account the integration time of the image sensor – in nJ/cm2.

The image sensor produces an output in Digital Numbers or “DN” and we get Responsivity in “DN/nJ/cm2”.

This is an accurate method when using narrow-band light sources or filters. In the Black Body  scenario – with broadband (“white”) light – one would set the optical power meter to some “average” or “peak sensitivity” wavelength (in this case ~700nm) to get reasonable readings.

How does this relate to errors in Responsivity?

In test 1 the system was used without the IR cut filter, so the whole spectrum of emission was available (Signal ~ input spectrum x QE). The optical power meter receives the full spectrum, applying the 700nm-equivalent gain. The red graph below shows the spectrum of acquired signal (electrons) for the sensor(s). The center of the spectrum sits around 700nm. With this setup both sensor and power meter will give a specific output value based on their respective spectral sensitivity and gain setup.

Photon Emission - Full Spectrum absorption

Figure 2: Electron Generation in Silicon – no IR cut

In test 2 the system is used with IR-cut filter. The energy impinging onto both sensors is much less (area under the curve) but – more importantly (Note) – the spectral content has shifted significantly towards the blue – the large deep-red portion is blocked.

Photon Emission - IRcut Spectrum

Figure 3: Electron Generation in Silicon – with IR cut

The center point of the Signal distribution has shifted towards ~600nm. As the light meter is less sensitive in blue a higher gain factor would be required to accurately represent the energy content in the spectrum. The differences in spectral response between meter and image sensor result in the impinging light energy being measured too low and the calculated Responsivity appearing too high.

In our accidental filter test, one could argue that we could simply have shifted the light power meter calibration value to 600nm, as that seems to be near the new center of the spectrum. While this may have resulted in a “closer” match it is still a distortion on the original measurement, since the difference in spectral response between the meter and the sensor has to be considered when the test setup is changed.

Using a single wavelength calibration with a given light source is acceptable in a production environment, when sensor-to-sensor variances are to be found, the setup never changes and no absolute data references are required.

In a careful measurement one must consider that a given spectrum cannot reliably be represented by a single wavelength, neither in the image sensor nor in the power meter. Therefore, to create “absolute” data we measure detailed Spectral Responsivity with spectrally controlled lighting (such as LEDs, Narrow Band Filters or Monochromators) and measurement equipment that is calibrated and carefully controlled.

Luckily, knowing these impacts one can distinguish between acceptable and unacceptable measurement errors. Using a single wavelength calibration with “white” light simplifies a production setup while Sensor R&D can utilize the advanced, more time consuming equipment and test methods.

Notice that I put “white” (light) in quotation marks… The second part of this series will discuss what “white” light can mean in: “How white are your Whites?”


Until Then,



Note: The spectral content – as opposed to the total energetic content – is more important only for this discussion. If one would do this test “live” and insert/remove the IR cut filter without changing the test arrangement the spectral change would be directly noticable in reduced sensor output signal (compare red curves in Fig. 2 vs. Fig. 3). In our circumstances we adjusted the light power to reach an acceptable output signal level and thus did not notice (immediately) the overall attenuation in signal due to the filter.


About Matthias

Born in the early seventies in Northern Germany I graduated in Physics at the University of Ulm. I worked for the BOSCH automotive group in Stuttgart before joining DALSA in 2000. Since then I have worked on layout, design, project lead, architecture and program management for various DALSA CMOS image sensors. I enjoy creating sensors and seeing them through production and implementation at the customer end.
Posted on by Matthias. This entry was posted in Machine Vision. Bookmark the permalink.

Comments are closed.