600 x 600 mm AOI: Field-of-View Budget and Stitch Count for CoWoS and FOPLP Panels

Uni-Ring publicly describes AOI platforms covering panels up to 600 × 600 mm with 2D AOI plus 3D Z-height measurement for CoWoS, FOPLP and CPO work. At that size the number that decides takt time is not camera resolution — it is how many fields the machine has to take, and whether the stage can put them back together inside tolerance. This article works that arithmetic out and shows where the budget actually goes.

The field of view, not the megapixel count, sets everything else

Field of view is sensor width divided by magnification. Once you know it, the sampling in micrometres per pixel, the number of fields across the panel and the total cycle time all follow. Choosing a camera first and then asking what it can see is the wrong order, because the panel does not change size to suit a sensor. Every number below starts from the field, and the wide-field optics that make a larger field usable are covered in choosing a telecentric lens: the four numbers that decide.

The stitch count, worked out

Take a 600 × 600 mm panel and square fields with 10 percent overlap. At a 20 mm field the usable step is 18 mm, so each axis needs 600 ÷ 18 = 33.3, which rounds up to 34 fields: 34 × 34 = 1,156 fields per panel. At a 50 mm field the step is 45 mm, giving 600 ÷ 45 = 13.3, rounded up to 14: 14 × 14 = 196 fields. Widening the field 2.5 times cuts the field count by nearly six times, because the count scales with the square of the field size. These are geometry calculations, not a vendor's takt time: they exclude acceleration and settling, per-field autofocus, and any re-measurement of failed fields.

Why a bigger field is not free

Widening the field at the same sensor means lowering the magnification, which means more micrometres per pixel, which means the smallest defect you can resolve grows. So the trade is asymmetric: field count falls with the square of the field size while detectable defect size grows linearly with it. There is a second cost that is easy to miss — a wide-field lens with the distortion budget that stitching needs is larger, heavier and more expensive, and its residual distortion becomes stitching error directly. For what the design buys you in the first place, see what a telecentric lens does and when you need one; a concrete example of a wide-field object-side telecentric build is the LTCM02-110 0.2X telecentric lens.

Overlap is the stitching datum, not padding

Overlap exists so that adjacent fields share features the registration algorithm can match. Too little and you get seams or gaps; too much and the field count climbs quickly. On the 20 mm example, going from 10 percent to 20 percent overlap moves the step from 18 mm to 16 mm, taking each axis from 34 fields to 38: 1,156 fields become 1,444, about 25 percent more motion and exposure time for the same panel. The other half of the seam budget is stage accuracy — stitching error is dominated by encoder repeatability and by how the panel is held, not by camera resolution.

Height adds a second axis of cost

If the recipe includes 3D Z measurement, that channel usually costs more time per field than the 2D capture, because height has to be scanned or projected rather than grabbed in a single exposure. Panel flatness then becomes a time problem: if the substrate is not flat to within the depth of field, every field needs its own autofocus, and autofocus multiplied by 1,156 fields is a real line item. This is why the fixture and the panel support belong in the budget discussion rather than in the commissioning punch list. Single-shot systems avoid the per-field focus question for small parts altogether — see how single-shot flash measurement works and when it fails and the IMAGE 3 instant vision measuring system.

A budgeting template you can fill in

Write down, in this order: panel size and usable area; the smallest defect you must catch, which fixes micrometres per pixel and therefore sensor, magnification and field of view; the overlap your stitching algorithm needs; fields per axis and total fields; time per field including exposure, move, settle, autofocus and any 3D scan; the expected re-measure rate; and the resulting takt against the line requirement. If the total misses the target, the levers are field size, overlap, per-field time and the number of heads — in that order of leverage. For continuous in-line work on this scale, a dedicated head such as the PMS-HM1120 telecentric measurement system moves the problem from stitching to sampling rate.

Frequently asked questions

How many fields does a 600 x 600 mm panel take at a 20 mm field of view?

With 10 percent overlap the step is 18 mm, so each axis needs 34 fields and the panel takes 34 × 34 = 1,156 fields. At 20 percent overlap the step drops to 16 mm and the panel takes 38 × 38 = 1,444 fields.

Does a larger field of view always make inspection faster?

It reduces the number of fields, which is usually the dominant term, but it also lowers magnification and raises micrometres per pixel, so the smallest detectable defect grows. It is faster only while the sampling still meets the defect specification.

Why do stitching algorithms need overlap at all?

Overlap gives adjacent fields common features that the registration step can match. Without shared features there is nothing to align to, and the seam error becomes the stage's positioning error rather than a measured quantity.

Is stage accuracy or camera resolution the limiting factor?

For a stitched panel, seam error is usually set by stage repeatability, encoder resolution, thermal drift and how flatly the panel is held. Camera resolution determines what each field can resolve, not how well the fields join.

Can a larger sensor replace stitching?

Only up to the point where one field covers the panel at the sampling the defect specification needs. At 600 mm and single-digit micrometre sampling, one field generally cannot cover the panel, so stitching remains.

Does panel warpage change the budget?

Yes. If the panel is not flat within the depth of field, each field needs autofocus, and that time multiplies by the number of fields. Warpage also makes a single global focus impossible, which is a fixture problem as much as an optical one.