.::APHESA SPRL - imaging, electronics and software consulting::.

Why you can't use a white (broad spectrum) light source for your spectral measurements

We are sometimes asked to perform fine spectral measurements of the performance of image sensors or cameras, something that our standard EMVA1288 equipment can't do because of its LED based design. LEDs are not really monochromatic as they have a band at 50% intensity varying between 10nm to more than 50nm.

Most labs will use a wide spectrum, usually white, light source and a monochromator to provide the spectral response curves. This approach can lead to incorrect or not precise results. We will explain why and present our approach.

Commonly used method

The common approach is to use a wide spectrum light source, for example a tungsten source, and filter the light source with a monochromator to only keep a narrow band spectrum centered around the test wavelength. The FWHM (Full Width Half Maximum) of such a spectrum is typically 5nm, 10nm, 20nm or more.

This beam then hits the sensor on a small area and the center of the spot, typically about 10 to 50% of its area, is used to measure the average sensor output for each color channel. After calibration, this average value represents the responsivity or the QE (Quantum Efficiency) of the sensor for that central wavelength.

The measurement is then repeated in steps equal to or smaller than the FWHM until the complete response range of the sensor is covered.

Most companies will then report the R, G, B curves or the monochrome curve of the sensor's response and claim that this is the sensor's response.

They are wrong

That statement is wrong! The R, G, B or the monochrome curve is not the response curve of the sensor but only some information about it.

First of all, it is usually only measured for an incident beam perpendicular to the sensor, or in other words at a zero chief ray angle. But we know that the spectral response varies in intensity and in shape with the angle of incidence. It also depends on the direction of the incidence as changing the angle in X (horizontal or along a row) does not give the same results as changing the angle in Y (vertical or along a column). This is highly related to the internal pixel structure, mostly metal layers, but also the technology and the design of the microlenses. Optical crosstalk and material absorption can also play a role and their effect is not symmetrical.

Secondly, it is a noisy measurement and therefore the precision depends on the number of samples taken. The measurement is noisy because the amount of light after the monochromator is limited as only a tiny fraction of the source's spectrum reaches the sensor. Therefore the signal level is small and its SNR is therefore low. In order to get more signal, it is possible to use more gain or a longer exposure time but this won't solve the SNR problem. In some cases, it could even make it worse. This is more extreme at the edge of the spectrum, both NUV (Near UltraViolet) and NIR (Near InfraRed), where the sensor's sensitivity is very low and a lot more signal would be required.

The shape of the spectrum also depends on the bandwidth of the measuring instrument. As the response spectrum can exhibit oscillations at several scales (see this other publication about spectral response), only a very small bandwidth, i.e. a narrow monochromator lid, can reproduce the actual shape of the response curve, any other approach will only produce a smoothed curve without any detail. This is very true for FSI (Front Side Illuminated) CMOS and less critical for BSI (Back Side Illuminated) sensors and CCD sensors. CCDs have a less complex structure that causes less oscillations and BSI sensors have the complexity of the metal structure after the photodiode and therefore this structure has a lot less influence. However, some level of details might be required for some applications, especially the applications that involve a laser or a luminescence or fluorescence phenomenon at a specific wavelength. It is also important for multispectral and hyperspectral applications for which each band is narrow and therefore high frequency oscillations or pixel to pixel variations of the spectrum are more critical.

Finally, the wide band light source will change over time and its spectrum will change over temperature, therefore requiring regular calibration, sometimes even calibration during the test itself. The light source may also flicker and this effect is more visible at shorter exposure times.

Therefore, these curves only make sense if the incidence angle, the slid's width (or similarly the bandwidth) of the monochromator, the temperature, the size of the measured spot and the exposure time and mentioned with the plot.

Our approach

To overcome these difficulties, we obviously need to provide the response curves at several incidence angles in both directions, but we also use a narrow bandwidth and a small number of pixels. This is obvious, but how can this be achieved?

Most of the problems come from light power. A higher incidence angle provides less light to the photodiodes, a narrower spectrum has less intensity, less pixels to average means a less good SNR. The solution is therefore to have more power from the light source but this is not easy to achieve, buying a stronger light source causes other difficulties.

The light source that we use is based on a femtosecond laser and a supercontinuum.

The femtosecond laser is a solid laser that provided very high power light pulses with a duration in the range of the femtosecond. Lasers are known to provide coherent, repeatable and high intensity light. Our laser is red and the laser has the size of a small table.

The role of the supercontinuum is to turn the monochromatic laser light into a wide spectrum by a collection of non-linear processes. The supercontinuum is a long microstructured optical fiber. The fiber seems red at its beginning and is white in the end as the spectrum broadens along the fiber.

With this equipment, a predictable and high intensity wide spectrum light is applied at the input of the monochromator. This allows for high SNR measurements for bandwidths of 10nm or more, even for a small number of pixels or with an incidence angle. With less incidence and some more pixels, acceptable results can be achieved for a bandwidth as low as 1nm. Using two separate fibers, we can cover the spectral range from 200nm up to 1650nm.

A typical measurement will provide the curves at 10nm bandwidth or more for the straight light and 9 other (X,Y) incidence combinations and a small area of interest. The measurement is typically repeated for the center of the sensor and two locations off center. We will then also provide as many curves as possible at 1nm bandwidth. The data is not provided only as curves but as a CSV table containing the details of the measurements so that numbers can easily be extracted from the graphs.

As for any other measurement, it is preferred to measure several sensors, ideally from several batches, and in a large enough quantity so that the extracted data can be used to calculate a statistical envelope of a typical device instead of the measurement of a single device.


About us

Aphesa develops custom cameras and custom electronics including FPGA code and embedded software. We also provide EMVA1288 test equipment and test services as well as consulting and training in machine vision and imaging technologies. Aphesa works in several markets including industrial, medical, oil&gas and security.