Dark Signal Non-Uniformity (DSNU) is a measure of the level of time-independent variation in the background of a camera’s image. It provides a rough numerical indication of the quality of that background image, with regards to patterns or structures that can sometimes be present.
In low-light imaging, a camera’s background quality can become an important factor. When no photons are incident on the camera, images acquired will typically not display pixel values of 0 grey levels (ADU). An ‘offset’ value is typically present, such as 100 grey levels, which the camera will display when no light is present, plus or minus the influence of noise on the measurement. However, without careful calibration and correction, there may be some variation from pixel to pixel in this fixed offset value. This variation is called ‘Fixed Pattern Noise’. DNSU represents the extent of this fixed pattern noise. It represents the standard deviation of the pixel offset values, measured in electrons.
For many low-light imaging cameras, DSNU is typically below around 0.5e-. This means that for medium- or high-light applications with hundreds or thousands of photons captured per pixel, this noise contribution is absolutely negligible. Indeed, for low light applications too, providing the DSNU is lower than the camera’s read noise (typically 1-3e-), this fixed pattern noise is unlikely to play a role in image quality.
However, DSNU isn’t a perfect representation of fixed pattern noise, as it fails to capture two important factors. Firstly, CMOS cameras can exhibit structured patterns in this offset variation, often in the form of columns of pixels differing from each other in their offset value. This ‘Fixed Pattern Column Noise’ noise is far more visible to our eye than unstructured noise, but this difference is not represented by the DSNU value. These column artefacts may appear in the background of very low light images, such as when the peak detected signal is less than 100 photo-electrons. Viewing a ‘bias’ image, the image the camera produces without light, will allow you to check for the presence of structured pattern noise.
Secondly, in some cases, structured variations in offset can be time-dependent, varying from one frame to the next. As DSNU shows only the time-independent variation, these are not included. Viewing a sequence of bias images will allow you to check for the presence of time-dependent structured pattern noise.
As noted however, DSNU and background offset variations will not be an important factor for medium- to high-light applications with thousands of photons per pixel, as these signals will be far stronger than the variations.