Object-Side Resolution vs Measurement Accuracy

Object-side resolution is the size of one camera pixel projected onto the part: sensor pixel pitch divided by magnification. A 1.85 µm pixel behind a 0.044× lens gives about 42 µm on the object. It is the smallest increment the image can represent — it is not the uncertainty of a measurement. The error of a real station is dominated by other terms, and a well-built system typically measures to roughly three to ten times the object-side pixel size for a defined feature under stable conditions. The multiple has to be established by measurement, not assumed from the pixel pitch.

Why a single edge can beat the pixel, and a measurement usually cannot

Edge position is interpolated to a fraction of a pixel — around one tenth of a pixel for a high-contrast edge with low noise — so the repeatability of one edge can be finer than the pixel. Accuracy is a different quantity: it includes how far the measured value sits from the true value, which brings in calibration uncertainty, distortion residuals, lighting-induced edge bias and mechanical stability. This is also why repeatability and accuracy should never be quoted as one number. Modern stations usually have good repeatability and substantially worse accuracy, and acceptance decisions under ISO 14253-1 have to account for the difference through a guard band.

What determines the achievable edge

• Contrast across the edge — a washed-out transition moves under noise, whatever the algorithm.
• Optical transition width, set by the lens MTF and the aperture: a sharper transition localises better.
• Noise, which sets the floor for interpolating a position.
• The algorithm: threshold, first-derivative peak and sub-pixel line fit each give a different bias, so the choice has to stay fixed inside a qualified recipe.

Building an error budget, and verifying it

List the independent contributions with a number for each: calibration artefact uncertainty, edge repeatability on the real feature, lighting bias, magnification error where heights vary (h / WD), distortion residual inside the used field, thermal drift over the working day (steel expands about 11.5 µm per metre per kelvin, aluminium about 23), stage or fixture positioning, and algorithm bias. Combine random terms in root-sum-square and systematic terms worst-case, then apply a guard band against the tolerance. Verify with a certified artefact measured at the centre, mid-field and edge of the field, at each height the part presents, repeated at least thirty times, and repeated again at the start and end of a shift.

Related reading: distortion, accuracy and cost compared, telecentric lens selection, and the telecentric lens range. The individual questions on resolution, repeatability, lighting and error budgets are answered below — send us the feature and the tolerance and we will run the arithmetic for your case.