Exposure Time, Gain and Motion Blur

Exposure time and trigger timing produce the same class of failure — an edge that is not where the algorithm assumes it is — but exposure is the easier one to reason about, because it is arithmetic with numbers you already know. The length of the smear is decided entirely by how far the part travels while the sensor is collecting light. Nothing about the algorithm, the lighting or the resolution changes it. Which means it can be calculated at selection time, from the line speed, and never has to be discovered on site.

The blur formula

Smear in pixels equals the object speed in millimetres per second, multiplied by the exposure time in seconds, divided by the object-space pixel size in millimetres per pixel. All three quantities come from the specification: the speed from the conveyor or the indexing table, the exposure from the camera setting, and the pixel size from the field of view divided by the sensor width. Take a line running at 300 mm per second with an exposure of 500 microseconds. The part travels 0.15 mm during the exposure. If the object-space pixel size is 0.05 mm per pixel, that 0.15 mm covers three pixels — the edge is no longer an edge but a gradient three pixels wide.

That number has to be compared against what the inspection actually needs, and this is where the conversation usually goes wrong. Sub-pixel edge detection can locate an edge to a quarter of a pixel or better on a sharp transition. Three pixels of smear does not merely add error to that, it removes the sharp transition that sub-pixel estimation depends on. The requirement is therefore not "blur smaller than a pixel" but "blur negligible compared with the sub-pixel accuracy claimed", which in practice means keeping the smear below about a quarter of a pixel for dimensional work.

Solve exposure first, then brightness

The right order is the opposite of the intuitive one. Work out from the speed and the required accuracy the longest exposure the task can tolerate, set that as a fixed ceiling, and only then ask how to get enough light within it. Brightness has several levers and they are not equivalent. Opening the aperture gathers more light without any cost in time, but it reduces depth of field, which matters when the part height varies. Increasing the source output is usually the cheapest and most reversible. Extending the exposure is not available, because the ceiling was set by the physics of the moving part.

Gain sits last in that list and it is routinely used first. Gain multiplies the signal and the noise equally, so it brightens the image without improving the signal-to-noise ratio at all; it moves the histogram to the right while making every grey level less certain. A station running at high gain will pass acceptance on a clean sample and then produce unstable grey levels on a slightly different surface. If a design needs significant gain to reach a usable brightness, the real problem is the light, and adding light is almost always cheaper than the consequences of the gain.

Sequential readout inverts the intuition

A sensor with sequential readout exposes and reads its rows one after another rather than all at once. The consequence is counter-intuitive: shortening the exposure does not eliminate distortion on a moving part, because while the rows are being read the part keeps moving, so the top and bottom of the image are captured at different positions. The part is sheared even though no individual row is blurred. For a part moving across the field of view this appears as an apparent skew; for a rotating part it appears as a warp. A sensor with simultaneous exposure avoids this entirely. Whether that matters depends on the part, but the test is decisive: image a part while the line is moving and look at a feature that should be a right angle.