Pick-and-Place Camera Aperture: Which Data Locks the f-Number

Ask the maker to quantify both, but judge the two numbers on one ledger: lock the widest aperture whose worst-case edge-localization error, expressed in object-space microns, still meets your tolerance across the entire height band the robot presents. Depth gained is the requirement. Resolution lost is the acceptance limit. What decides the lock is not a single figure but an aperture sweep that reports usable object-space depth and edge error together, at the near, nominal and far object planes, at your magnification, your wavelength and your illumination. A depth figure without a matching error figure cannot tell you which f-number to freeze, and a complaint about soft edges cannot tell the maker what to change. Numbers first, impressions second: a blurred edge is a measurable quantity, and it is the only version of the problem a lens house can act on.

For factory-automation builds, the Lensmind FA lens series lists the large-format models with sub-0.1% distortion grades.

Why Fine Edges Soften When the Iris Closes

Stopping down did what you asked. A larger f-number narrows the cone of light the lens accepts, which stretches the range over which a point on the part still lands inside an acceptable blur circle on the sensor. That same narrowing spreads each point into a wider diffraction pattern, and contrast at fine spatial frequencies falls with it. Past a certain setting, the limit is no longer the polish on the glass but the diffraction limit of the aperture itself.

The response to aperture is a curve with a bottom rather than a slope. Wide open, the aberration terms dominate: spherical aberration, coma, field curvature and lateral colour all inflate the blur spot, and stopping down shrinks them, so edges often sharpen through the first two or three stops. Beyond the minimum, diffraction grows and edges soften again. The best aperture for your measurement sits at the bottom of that curve, not at either end of it.

That bottom also moves across the field. A lens that is sharpest at f/4 in the centre may bottom out at f/5.6 or f/8 near the corner, because off-axis aberrations fall faster with aperture than diffraction rises with it. If your edge measurement runs anywhere in the outer field, the worst point decides, not the centre. Field of view and image circle decisions made before the drawing was released come back later as an aperture constraint.

What to Ask the Maker to Quantify: Two Curves and a Sweep

Ask for two measured curves rather than an opinion about image quality. The first is usable object-space depth against f-number, taken at your working distance and magnification, and defined against a stated blur criterion instead of a visual judgment about focus. That criterion has to come from your tolerance: the acceptable circle of confusion is the object-space error you can absorb multiplied by the magnification, so the maker cannot draw the curve properly until you hand over the smallest edge you measure and the position error you allow on it.

The second curve is contrast at a spatial frequency against f-number, measured at the frequency that corresponds to your feature, at the worst field point, at your wavelength and under your illumination. A generic trace at 50 lp/mm describes a lens, not your edge. If the maker has data at only wide open and one middling stop, the shape between the points is missing, and the shape is the decision. Those traces are easier to use once the axis labels and the on-axis and off-axis conventions are clear, and the guide to reading an MTF chart for a machine vision lens walks through them.

Then ask for the sweep itself: the same two quantities at several apertures, at three object planes (nearest part surface, nominal, farthest), with illumination, exposure and gain recorded on every row. Hold those three steady across the sweep if you can, because closing the iris costs light in proportion to the square of the f-number, and buying the exposure back with sensor gain or a longer integration adds noise and smear. Those are measurement terms, not photography terms: noise moves a sub-pixel edge fit even when the optics have not changed at all.

The Working f-Number Your Datasheet Does Not Print

Published depth and resolution numbers usually describe a lens focused near infinity at its marked f-number. Close-range gauging does not behave that way. At magnification m, the working f-number is approximately the marked value multiplied by (1 + m) for a symmetric lens whose pupil sits near the principal plane. Set the ring to f/8 at 1x and the aperture behaves, for diffraction and for light gathering, like f/16. That factor is why a lens can look excellent on a bench at infinity and deliver soft edges the moment it is racked out to work 100 mm from a part.

The blur scales the same way. In object space, the diffraction spot is roughly 2.44 times the wavelength times the working f-number, divided by the magnification. With 500 nm light, a marked f/8 at 1x gives a working f-number near 16 and an object-space diffraction blur near 20 microns, which is a large fraction of any edge you would care to locate at that magnification. Depth at the same setting follows the familiar close-range form: about twice the marked f-number, times the acceptable image-space blur, times (1 + m), divided by m squared. Both quantities carry the working-aperture factor, and they carry it in the same direction.

The practical consequence is a request rather than a complaint. Put the magnification, or the field of view and the working distance, into the data request, and require every depth and contrast curve to state the working f-number and the wavelength it was measured at. A curve without those two labels cannot be compared against anything you will see on the machine, and it cannot be compared against a second supplier’s curve either.

Why No Aperture Can Change the Ratio You Are Trading

Divide the object-space diffraction blur by the object-space depth at the same setting and magnification and the f-number cancels out. What remains is a constant: roughly 1.22 times the wavelength, divided by the object-space error you are prepared to accept. Work the numbers at 500 nm with a 5 micron error budget and the depth you can cover comes out at about eight times the blur you create. That ratio is identical at f/4 and at f/11. Stopping down slides you along a line of constant trade; it cannot move the line.

This is worth having in mind before you argue with a supplier, because it changes what you ask for. If the ratio is fixed by wavelength and by your error budget, the only ways to buy depth without paying for it in blur are to shorten the wavelength (blue or near-UV light, where the part and the coating allow it), to widen the error your algorithm will tolerate (calibration, a model that fits the real edge profile, more pixels across the edge), or to reduce the depth you actually need. The last option is the cheapest and the one most often skipped.

Splitting the depth requirement is where that reduction comes from. Systematic height offset, the fixed difference between where the parts sit and where the lens is focused, can be shimmed, re-focused or designed out of the fixture. Only the random part has to be covered by depth of field: part-to-part thickness variation, placement runout, and any tilt that walks across the field. Size the depth against the random portion and treat the systematic portion as an alignment task. A longer treatment of the specification side sits in the site’s guide to specifying depth of field without guessing.

How the Sweep Is Run and What Locks the Ring

Run the sweep on the machine, or on a bench that reproduces the machine, with the illumination and the fixture that will run in production. Every row of the record needs the same columns, and they should be filled by measurement rather than by calculation: aperture; usable object-space depth at the stated blur criterion; contrast at your feature frequency at the worst field point; the repeatability and the bias of your own edge measurement on a known artifact; and the exposure and gain used to capture the frame. Bias deserves more attention than repeatability, because a blurred edge can be located with very little scatter while sitting a consistent fraction of a pixel away from the true position, and a steady offset will not show up in a scatter number.

What you are looking for in those rows is the widest aperture whose worst-case plane, at the worst field point, still passes. Depth beyond that requirement buys nothing and costs contrast, so the target is not the largest figure in the depth column. The two objects that matter most for the decision are the ones at the ends of the height band, so an acceptance test that checks only a flat gauge at nominal height will not protect you.

Then lock the ring and document the state. Mechanical locking is not a formality: M12 S-mount lenses on embedded rigs use thread-locked focus for exactly this reason, and an aperture ring with a set screw costs less than one lost shift of scrap. The as-installed record should carry the aperture, the working distance, the magnification, the wavelength and illumination, the exposure, the gain, and the flange arrangement. That last field is easy to forget and expensive to omit: a different mount or spacer changes the flange distance, and converting C-mount to CS-mount through a 5 mm ring is a magnification change, which moves both curves. C-mount sits at 17.526 mm and CS-mount at 12.5 mm, so the adapter is not cosmetic.

It also pays to ask for the numbers on the batch you will actually run. Lensmind ships test documentation with every production batch covering the agreed family parameters, typically MTF, distortion, BFL and DOF, with the batch number, the test conditions, the inspection coverage and an explicit pass or fail judgment, as a default rather than on request. The asking is about labels: which aperture, which magnification, which wavelength, which field point. With those attached, incoming inspection has something to compare against, and a lens that drifts on re-order shows up in the document instead of in the yield.

When No Aperture Wins, Change the Optics Instead

If the sweep shows that no setting meets the requirement, the answer is not a different f-number. It is a different optical arrangement.

Telecentric designs are the standard fix when measured dimensions have to stay constant with depth. An object-side or bi-telecentric lens holds magnification across a depth range, so defocus appears as blur rather than as a scale error, which is often the failure that made you close the iris in the first place. The trade is field size and front element cost, and both scale together.

A wider field at lower magnification with more pixels is another route, and the sampling gate is worth checking before committing to it: the usual machine vision rule wants three to five sensor pixels across the smallest feature you must locate, which in object-space terms means the pixel footprint at the part, sensor pitch divided by magnification, has to be a third to a fifth of that feature. Diffraction has a matching gate. Once the blur spot grows past roughly two pixel pitches, resolution added on the sensor side no longer buys accuracy.

A liquid lens changes the argument instead of the aperture. The ZWA series (M0834AFC, M1238AFC, M1643AFC, H1670AFC-A, H1670AFC-B) shifts focus by electrowetting in about 20 ms with no moving parts, which makes per-plane refocusing or a short focus stack practical inside a cycle rather than asking one aperture to swallow the whole height band. That route spends cycle time or software complexity to recover the depth the iris could not deliver cleanly.

The aberration term belongs on the table as well, because it sets where the bottom of the aperture curve sits. A lens that is well centred, with element spacing and centring held to the micron class, can run wider before blur takes over, and the aperture where depth is cheapest is the widest one that still holds tolerance. Alignment and metrology work of that kind, 1 micron class active alignment plus centration and surface measurement on instruments such as Trioptics and a laser Fizeau interferometer, moves the bottom of the curve rather than the ring position.

For close-distance gauging, the FA C-mount catalogue covers 2/3 inch to 1.1 inch formats with distortion held below 0.05% and 10K+ batch consistency, and includes a macro group for short working distances, with models such as FA3505W, FA5005W, FA2520W, FA3520W, FA5020W and FA7520W alongside the LMHS-MH series, MH1825, MH2012, MH2025 and MH2406 among them. When the requirement falls outside the catalogue, a drawing or a sample can carry a custom design through to production, quoted at 60 working days from prototype to mass production, with a first-pass answer on a custom enquiry inside 48 hours.

Write the requirement as one sentence and the rest of the discussion follows from it: usable object-space depth of X microns at aperture Y, with edge-position repeatability of Z microns and contrast at your feature frequency at or above a stated value at the worst field point, at magnification M, wavelength L and working distance D. Any sweep the maker returns can then be scored against it line by line, and the aperture you lock will be the one that sentence picked, not the one that looked sharpest on a monitor.