MVLscope LESSON 06 / About 32 min
Follow a photon into a number.
Detectors, Noise & Exposure
Build an electron budget, separate noise sources, and discover why a brighter camera image can contain no more evidence.
Before you begin
Illumination, Contrast & Exposure · Sampling & Pixels
Try a setting, save an observation, then change one thing. The explanation checks are optional.
Loading the interactive experiment… If it does not appear, reload the page. The lesson outline and model notes below remain available.
MAKE SENSE OF WHAT YOU SEE
A QUESTION TO THINK ABOUT
See the assembly in 3D Optional · all controls also work above
Optional 3D view. The calculated diagram above is the measurement reference.
YOUR EVIDENCE
Keep the comparison.
Saved images and readouts stay as they were when you captured them. Save a second setting to compare side by side.
Your notebook is empty. Use “Save this observation” beside an experiment.
Lesson outline
A photon is an opportunity
A pixel collects light over an area and a time. Quantum efficiency (QE) is the probability that an incident photon contributes a detected photoelectron in this simple visible-light camera model.
An image starts as a collection process. Count the expected electrons before judging its brightness.
Noise sources add as variances
Noise is the spread of repeated measurements. Photon shot noise, background shot noise, dark-current noise and read noise have different origins.
Reducing the dominant variance helps most. SNR and brightness describe different properties.
Two different ceilings
A pixel has a charge capacity, and its digitizer has a largest code. The earlier ceiling determines when distinctions are lost.
Check both charge capacity and ADC headroom. Reducing display brightness cannot recover clipped measurements.
Same brightness, different evidence
Digital units, electrons and display brightness are three different scales. Compare a short brightened acquisition with a longer reference acquisition.
Match acquisition units before comparing cameras. Matching the appearance of two images does not match their evidence.
One long read or many short reads?
Hold the total integration time fixed and divide it among several frames. Summing those frames preserves the expected signal but changes the read-noise cost and per-frame headroom.
Choose the number of reads using both noise and headroom, while accounting for the experiment’s timing.
Model, limits & sources
Teaching model for a linear visible-light CCD/CMOS pixel; QE is constant at a chosen wavelength. No EM gain, avalanche multiplication, pixel-response nonuniformity, correlated noise, blooming, hot pixels or rolling shutter.
SNR describes expected, unsaturated signal with an exactly known mean background. Image pixels are independent Poisson photoelectrons plus Gaussian read noise, followed by full-well clipping and ADC rounding. A zero-bias ADC clips negative readout values, which can bias the faintest pixels upward.
The colored bead pattern is a fixed synthetic sensor-plane phantom, not a resolution simulation or biological specimen. Red marks acquired camera clipping. The seeded frame is reproducible, not a confidence interval.
Conversion gain is ADU/e⁻, the reciprocal of the e⁻/ADU convention used by some manufacturers. Display gain changes rendering only. Quantization error is not included in the main SNR and its RMS approximation is shown separately.
Read further
PUT THE IDEA TO WORK
QE depends on wavelength, and equal optical power need not mean equal photon arrival rates.