Backlighting reverses the usual imaging problem. Instead of fighting to make a feature visible against a variable surface, you place a diffuse light source behind the part and let the part block it. The result is a silhouette with extremely high contrast: part pixels near black, background pixels near saturation, and an edge that is a real physical boundary rather than a gradual shading transition. For dimensional work on a single plane, nothing else approaches it for stability.
What backlighting measures well
Hole diameter and position. A drilled, punched or moulded hole appears as a clean circular void. Sub-pixel edge detection on a saturated-to-dark transition routinely achieves repeatability of a few hundredths of a pixel, which on a 20 um per pixel setup is well under a micrometre of apparent resolution - far better than the optics actually justify, but very repeatable.
Gaps and clearances. Piston-to-bore clearance, connector insertion depth, seal gaps and thread pitch all read directly as light/dark widths.
Presence and count. Pins, teeth, slots and holes are counted by connected-component analysis on a near-binary image, which is about the most robust inspection algorithm that exists.
Edge position and straightness. Sheet edges, film edges and cut lengths are measured by tracking an edge position along the image, giving you both the dimension and its variation.
Opaque contamination in transparent media. Particles and bubbles in glass or liquid are trivially visible when backlit.
The right light source
Area backlights are flat panels of diffused LED, sized larger than the field of view so that the part sits wholly within the illuminated region. Uniformity is critical: any brightness gradient across the panel becomes a threshold that shifts with position, and a fixed threshold then measures different edge positions in different parts of the frame. Either specify good uniformity or calibrate the threshold per region.
Telecentric backlights are a specialty option for measurement at tight tolerance, combining a collimated backlight with a telecentric lens so that the silhouette has no parallax at all - the edge position is then truly independent of part height. They cost more than an ordinary backlight by a wide margin and are justified only when the tolerance demands it.
The limits you cannot design around
Backlighting gives you a silhouette, and a silhouette says nothing about what is on the part's surface. A dent, a scratch, a print error or a colour variation is invisible. For parts that need both dimensional and surface inspection, the usual answer is two stations with different illumination rather than one clever station.
Depth is the second limit. A backlit image is a projection. A tapered hole and a straight hole of the same entry diameter produce identical silhouettes. If the measurement depends on depth or shape along the optical axis, backlighting cannot resolve it at all.
Finally, backlighting requires optical access from both sides. That is trivial for a part presented on a clear conveyor or a gap in a transport rail, and impossible for a part mounted against a solid fixture or inside an assembly.
A practical measuring station
A backlit dimensional station is deliberately simple. Pick the resolution from the tolerance: a 1.3 MP sensor such as the MV-CS013-80GM across a 40 mm field yields roughly 30 um per pixel, adequate for a 0.1 mm tolerance with three pixels of edge margin. Add a low-distortion fixed lens - the MVL-HF1228M-6MPE is a reasonable 12 mm match for a short working distance - and a backlight panel an inch larger than the field in each direction. Calibrate once with a certified target, verify with a gauge pin, and the station will hold its accuracy for years because there is nothing in it that drifts.
Used within its limits, backlighting is the cheapest route to metrology-grade repeatability in machine vision. Used outside them, it is a system that confidently measures the wrong thing.