When characterizing the noise of an imager (temporal or fixed-pattern) by means of a single number (that is what we did in the previous blogs), one should be cautious with the interpretation of that single result. A single number does not tell anything about the distribution of the noise, does not give any information about outliers, does not give any information about the physical location of pixels with a high or low noise level. Performing a global noise measurement and characterizing the noise behavior of an imager by means of a single number can still work well in the case the noise is dominated by a global noise source, e.g. the output amplifier of a CCD. On the other hand, a single number for the noise is very often found in data sheet, but actually it can hide a lot of information.
But these days with amplifiers in every CMOS pixel, addressing circuitry on row level and analog circuitry on column level, it is absolutely recommended to perform more noise analysis to get a better insight in the origins of the noise. The following techniques can be of interest :
– Generating a noise distribution histogram if the noise is present on pixel level. In the case that the average noise level is dominated by outliers, it may be worthwhile to correct these outliers (by menas of a simple defect correction algorithm) and perform the noise calculation again,
– Calculating the noise on row level and on column level : this can give information about the origin of the noise, whether the dominant noise source is located in the column circuitry or row circuitry,
– Performing frequency analysis (by means of a FFT) on the row level and column level noise to find out if there exists any repetitive pattern in these noise sources,
– Calculating the noise on pixel level, by subtracting the column and row noise from the total noise,
– Calculating the minimum quantisation noise of the ADC (if present on the chip),
– Measuring the noise of the “global” output stage by applying different gains to this output amplifier (if possible).
The technique of noise measurements is not that difficult. With a computer and a frame grabber one can do an excellent job. In most cases it is just a matter of calculating average values and standard deviations on the data present in images. The trick is to perform them in the right order !
Over the last decade the noise performance of the imagers has increased quite a lot, for instance due to semiconductor technology improvements and advanced circuitry/design techniques. As a result, the characterization of the noise with a high accuracy is becoming more and more challenging. For that reason one should pay more and more attention to the measurement set-up as well. A few hints :
– Temperature of the sensor (which is not necessarily the same as the environment temperature) should be stable within 1oC. Even the smallest temperature increase will increase the dark current and its shot noise component,
– In the case that light is used as an input signal, the light source should be stable (intensity and colour temperature) and a DC light source has to be used. Uniformity of the light across the pixels should be “perfect” when various pixels will be compared with each other, as for instance done with PRNU measurements. If the case noise measurements are performend, one is should measure non-uniformities of the sensor and not non-uniformities of the light source.
– Measurements without light need to be done in dark ! Just avoiding light input by simply capping the camera lens is not enough to put a sensor in complete darkness.
Good luck with the measurements and have fun !