Build1 publisher3 min readPublished
Whether that tolerance means 0.03mm or 0.06mm sets the camera at 98MP or 65MP
An industrial vision guide puts pixel accuracy at one-eighth of the inspection tolerance or finer before any algorithm tuning is worth attempting. Its own 65MP worked example clears that bar only if a plus-or-minus band counts whole.
The Engineer · Build desk
What happened
- A dev.to troubleshooting guide for industrial inspection defines pixel accuracy as field of view divided by camera resolution and sets the safe threshold at one-eighth of the inspection tolerance or finer.
- That system finished with 0.01mm measurement repeatability, and the post argues that without enough resolution small defects are invisible in the image regardless of the algorithm.
- Rolling shutter exposes row by row with a delay between rows, so fast-moving workpieces come out tilted or stretched, while global shutter exposes every pixel at once.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Resolution and shutter type are chosen once, at purchase, so the one-eighth check fixes a project's accuracy ceiling before anyone opens a tuning ticket.
- cost Accuracy bought with pixels is paid for in frame rate and processing time, so on a fast line the integrator has to shrink the field of view or accept a looser accuracy target.
- exposure An acceptance test run on stationary or slow samples will pass a rolling-shutter camera that distorts every measurement once the line runs at production speed.
- constraint Nothing in the rule fixes the convention for a plus-or-minus figure, so a specification that cites one-eighth of tolerance cannot be verified until it says which reading it means.
Check the worked example against the rule and the answer turns on what "tolerance" means. The post's own illustration of the rule takes a stated 0.1mm tolerance and divides by eight to get 0.0125mm per pixel [2]. Apply that treatment to the connector job's plus-or-minus 0.03mm and the threshold is 0.00375mm per pixel [17]. The camera delivered 0.004597mm per pixel [3], about 23 percent coarser than that [18]. Read the plus-or-minus as a 0.06mm band and the threshold becomes 0.0075mm per pixel, which the same camera beats by 39 percent [19]. The post calls the result well above the requirement [3], and under the full-band reading it is.
The stricter reading is a big step up. Covering 42.95mm at 0.00375mm per pixel needs 11,453 pixels of width, and the 32.18mm height needs 8,581, which is roughly 98 megapixels against the 65 the team bought [20].
The eight in the rule is headroom for everything the sensor does not control. Repeatability on that job came in at 0.01mm [4] with pixels of 0.004597mm, so the shipped figure was about 2.2 pixels [21]. Edge extraction, part presentation, lens geometry and thermal movement all sit in that gap. Sub-pixel methods recover some of it, though none of them recover a defect that never landed on a pixel [6].
The shutter check costs one image. Move a calibration target at constant speed, photograph it, and look for straight lines that have gone slanted or curved [10]. Rolling shutter reads row by row with a small delay between rows, so a part moving during readout comes out tilted or stretched [8]. Hikrobot, Daheng Image and Basler all ship mature global shutter CMOS lines, according to the post, with Sony's Pregius and Onsemi's PYTHON sensors named as typical [11]. The piece also carries a correction note retiring its own CCD-versus-CMOS section, on the grounds that CMOS holds over 90 percent of the industrial camera market and CCD has largely left industrial vision [7].
Two symptoms open the post: a system that works in the lab and fails on the line, and equipment that passes acceptance and then spikes its miss rate three months later [15]. The post does not say how often the physical layer turns out to be the cause [24]. Of the three physical-layer directions published in part one, only calibration is a property that drifts after installation [25]. Insufficient resolution and a rolling shutter are wrong from the first image, and they survive acceptance when acceptance is run on stationary or slow-moving parts. For calibration the post lists the target's own accuracy, with ceramic more thermally stable than glass, along with ambient light during calibration and the number and angles of calibration images [14]. It quotes a workshop saying: "Poor calibration, all effort wasted." [13]
What to watch
- Part 2 of the guide, covering the algorithm-layer directions, which has not been published alongside part 1.
- Whether a revision states which convention the one-eighth rule uses for a plus-or-minus tolerance figure.
- A published takt-time figure for the connector housing line, which would show whether a higher-resolution sensor was ever feasible there.