A wrinkle in the adhesive layer under clear PET film on glass is nearly invisible in an ordinary transmitted-light image. The layer passes almost all the light whether it lies flat or ridged, so the defect hides in a few gray counts, and inspectors end up working in a dark room, hunting faint interference bands by eye. Judgments differ between shifts and cannot be audited. The fix is optical, not algorithmic: crossed polarized lighting, with one polarizing film on the backlight and a crossed analyzer on the camera, converts strain and thickness variation in the transparent layer into dark-field fringes a camera can threshold. This guide gives the physics, the parts recipe and the pixel math for a 200 x 350 mm laminated panel.
Why plain transmitted light misses the wrinkle
For a clear stack, image brightness is set by reflection at each surface and by absorption in the layers. PET has a refractive index near 1.64, so each polymer surface reflects ((1.64 - 1) / (1.64 + 1))² ≈ 0.059, about 6 %, and the glass adds a similar share. Those losses are identical on a good panel and on a wrinkled one, because the adhesive transmits light whether it lies flat or ridged. A wrinkle moves transmitted brightness by far less than the 1-in-256 step of one 8-bit gray level, which buries the defect in sensor noise.
What the camera actually sees is a bright field plus glare skipping off the two film surfaces and two glass surfaces into the lens. No threshold or classifier recovers contrast the sensor never recorded, so the lighting geometry has to create the contrast — the principle behind lighting geometry for image contrast.
What crossed polarizers do to a strained adhesive layer
Put one polarizing film under the panel and a second on the camera with its transmission axis crossed roughly 90° to the first. Light that keeps its polarization is blocked, so an ideal, unstressed stack images black. Stress and thickness gradients make the adhesive birefringent: the refractive index differs along the two principal directions, the two polarization components travel at different speeds, and they emerge with a phase retardation δ. Retardation twists the polarization, and part of the light then passes the crossed analyzer.
Δn is the birefringence, t the layer thickness, λ the wavelength and θ the angle between the polarizer axis and the principal stress direction. Brightness peaks where retardation is large and the stress axis sits near 45° to the film, so a wrinkle ridge draws itself as bands of equal retardation — position, width, curvature and area, all directly measurable. The mechanism is the one that shows Newton's rings between imperfectly contacting glass surfaces, and the same reason stretched plastic wrap glows between two crossed polarizing filters.
The parts recipe and the camera geometry
Four parts make the station: high-power diffuse backlights sized to the field, the two polarizing films, a monochrome camera over a red or green light — chosen by the same contrast logic as any LED colour pick — and a dark enclosure, because the camera-side film passes about half of any stray room light straight to the sensor.
For a 200 x 350 mm panel the lit area must cover the field in one shot. The LS-BKL-0865 is a high-power backlight with a 400 x 300 mm lit area, rated 92.2 W at 24 V, and its type list includes strobing. It covers this panel with margin on every side, and the power reserve matters: two films cost roughly two stops before the layer adds its own losses, as the arithmetic in this polarizer light-loss guide shows.
- Size
- 400 mm +
- Type
- High power / strobe
- Voltage
- 24 V
Mount the polarizer film on the light's output face and the analyzer in a rotatable filter holder on the lens. Keep the panel glass-side down, a few millimetres above the light. Keep the analyzer adjustable: turning it a degree or two off exact crossing lifts the whole field uniformly, which helps when the adhesive retardation is small. The camera then sees a field of view equal to the panel plus margin, the working distance follows from the focal length, and the next section sizes both.
Worked example: pixels, lens and exposure for a 200 x 350 mm laminated panel
The inputs, and where each comes from:
- Panel: 200 x 350 mm, the part this station screens.
- Field of view: 360 x 210 mm — panel plus 5 mm handling margin per side.
- Smallest band to resolve: 0.5 mm — a starting target for wrinkle fringes; substitute the number you measure on a known-bad panel, and every result below scales linearly with it.
- Detection rule: at least 3 pixels across the smallest band, the standard minimum for a stable threshold.
- Sensor: a typical 5 MP monochrome sensor, 2448 x 2048 pixels at 3.45 µm, an 8.45 mm wide active area.
The rule: pixel size must stay below band width divided by 3.
The arithmetic: 0.5 mm ÷ 3 px = 0.167 mm/px allowed. On the long axis, 360 mm at one pixel per sixth of a millimetre needs 360 x 6 = 2160 px; the short axis needs 210 x 6 = 1260 px. The 2448 x 2048 sensor clears both. Its actual sampling is 360 mm ÷ 2448 px = 0.147 mm/px, so a 0.5 mm band spans 0.5 ÷ 0.147 = 3.4 pixels, above the minimum with a small margin.
| Band target | Sampling needed | Pixels across 360 mm | Camera decision |
|---|---|---|---|
| 0.5 mm | 0.167 mm/px | 2160 | 5 MP (2448 px) fits |
| 0.25 mm | 0.083 mm/px | 4320 | 12 MP (4096 px) falls about 5 % short; 25 MP (5120 px) fits |
| 0.1 mm | 0.033 mm/px | 10,800 | beyond area scan; move to line scan |
The lens: magnification m = sensor width / field of view = 8.45 mm ÷ 360 mm = 0.0235. From f = WD x m / (1 + m), a 300 mm working distance needs f = 300 x 0.0235 ÷ 1.0235 = 6.9 mm. The nearest standard focal length, 8 mm, lands at WD = 8 x 1.0235 ÷ 0.0235 = 349 mm — enough clearance for the film holders. A 12 mm lens moves the camera out to about 520 mm if the station needs the room.
The exposure: crossed films put the clean field near black by design, so expose for the fringes. Start at f/4 and 10 ms with the light on continuously, and raise exposure until clean areas sit at 10-30 counts in 8-bit while the strongest band clears 100 counts. If bands stay under about 50 counts, drive the backlight in strobe mode and work up toward 20-50 ms before adding camera gain, because gain lifts noise with the signal; the timing limits are in this strobe lighting guide.
The decision: a 5 MP camera with an 8 mm lens at f/4, working about 349 mm above the panel, resolves 0.5 mm fringe bands across the full 200 x 350 mm field. Tighten the band target only by stepping up the sensor class, not the lens.
Diffuse or collimated light under the films
A diffuse flat backlight gives the most even field over a large area, which is what full-panel screening wants. Collimated lights send parallel rays through the layer, so the recorded image tracks the retardation map more sharply, and they pair naturally with telecentric lenses that hold perspective flat across the part — the trade-offs are compared in this collimated versus diffuse backlight guide, with the lens side in when a telecentric lens pays off for gauging.
The limit is size: the largest stocked collimated series prints beam diameters of 26 to 240 mm, so a full 350 mm panel under collimated light means scanning the part under one unit or tiling several. For coupons and small laminated parts, though, one unit covers the whole part — exactly what the bench test below needs.
The LS-BKL-0830 is a 120 x 120 mm collimated back light at 24 V, the LS-CLM-0079 projects a 90 mm beam from a 223.9 mm body at 3 V, and the LS-CLM-0063 is a Ø100 mm telecentric collimated light at 5 V with a 3-pin flat connector. Any of them covers a coupon large enough to carry your worst known wrinkle.
- Size
- 100-200 mm
- Type
- Collimated
- Size
- 50-100 mm
- Voltage
- 3 V
- Size
- 100-200 mm
- Voltage
- 5 V
Colour affects fringe countability: white light lays down rainbow-ordered fringes, while a single colour keeps them single-valued. For coupon tests, the LS-BKL-1468 red flat backlight covers a 150 x 150 mm lit area at 24 V, paired with the same two films.
From fringes to a pass/fail decision
Capture a reference stack of 20-30 known-good panels and average it. PET base film is stretched at manufacture and birefringent before anything is laminated on, so good panels show a stable baseline pattern, never black. Gate each new panel on deviation from that reference — pixel-wise difference or local fringe curvature — and raise the alarm at three times the good-part standard deviation, the 3-sigma starting rule, then tune it on your golden parts. Route flagged panels to a review queue instead of scrapping them; screening plus review beats the dark room quickly and makes every judgment auditable.
Judge thresholds on the live image rather than on saved samples, as the live binary image preview workflow does, and fix the optics before blaming the algorithm — the discipline in this glass bottle case study is the same. Once the optics hold steady, this walkthrough of commissioning a vision inspection for defect detection carries the software side: training sets, false-accept budgets and line sign-off.
What this lighting will not catch
Three limits to plan around. Particles with a refractive index close to the adhesive leave both transmittance and retardation nearly unchanged; catch them with the analyzer turned parallel — plain bright-field transmission from the same station — where they appear as dark points. Retardation wraps: a deep wrinkle that pushes δ past one full wavelength repeats the brightness of a shallow one, so count bands inward from a boundary or step through two wavelengths, as in this multi-wavelength sequencing guide. And the geometry reads the adhesive layer, not the glass: scratches and chips need reflective geometries, sorted in this surface versus silhouette lighting overview. For the wider family of transparent-material checks, this guide to selecting vision inspection for transparent polycarbonate and acrylic applies the same logic to other clear parts.
Checks before you order
Work down this list before spending; the first item decides everything after it.
- Hand test: sandwich a known-bad panel between two polarizing films and rotate them under a desk lamp. Bands at some angle confirm the defect carries usable retardation; no bands at any angle means transmitted lighting cannot show this defect, and no backlight will change that — move to multi-wavelength phase methods or keep review in the loop.
- Coupon on collimated light: repeat the hand test over a collimated unit and note the analyzer angle that gives the strongest band contrast; that angle becomes your production setting.
- Camera: re-run the pixel math with the band target measured from your worst part, keeping at least 3 pixels per band plus 10 % margin.
- Light: size the lit area to the field with margin on every side; pick red or green for the production recipe and keep white for debugging.
- Drive: confirm the light accepts a strobe channel and that the duty cycle holds at your cycle rate before committing to overdrive.
- Films: fix both holders rigidly, because a drifting analyzer angle drifts the baseline; if the films must live sealed and pre-crossed in one housing, that is a custom build — the special lights category includes a customized light series with 41 models.
If the hand test shows bands, the distance from the dark room to an auditable screen is one backlight, two films and one camera. Start in the light finder, check the camera against the calculators above, or step straight into the lighting configurator. The test plan for lens and lighting selection structures the bench work, and the lighting topic page holds the rest of the lighting guides for the questions this one raises.
Questions
Does this work with a colour camera or does it have to be monochrome?
Both work. A colour camera under white light records coloured fringe orders, useful when debugging because the colours rank retardation from violet to red. For an automated gate, a monochrome camera over a red or green light is simpler: fringes stay single-colour, thresholds hold across the field, and red usually gives the strongest sensor response. Change colour only when you need to count fringe orders.
How much brightness do the two polarizing films cost me?
Each linear polarizing film passes roughly 40-45 % of unpolarized light — about one stop per film on typical film data sheets — so the pair costs about two stops before the layer's own rotation decides what reaches the sensor. A crossed dark field is meant to have a near-black background, so expose for the fringes rather than the background. If they stay dim, run the light in strobe mode or open the lens one stop.
Do the polarizing films themselves cause fringes?
Yes, if they are loose. A sagging or creased film is itself stressed and birefringent, so it lays its own bands into every image and can mimic defects. Stretch each sheet in a rigid frame or sandwich it flat between glass, keep it a few millimetres from the panel, and capture a film-only dark frame during set-up to confirm the holders are clean. Replace films that get creased or scratched.
Will the same rig catch bubbles and dust in the adhesive?
Bubbles usually break the fringes themselves because they locally remove the birefringent layer — they show in the polarized recipe as fringe discontinuities. Particles that barely change retardation are a job for the parallel-analyzer bright-field recipe. Run both as separate gates with separate thresholds rather than one combined decision, so a miss in one recipe cannot hide behind the other.
What panel rate can one camera handle?
With a 5 MP camera and a strobed backlight, capture plus evaluation lands near one to two seconds per panel including handling. At 5,000 panels a day that is 10,000 seconds, about three hours of camera time, so one station covers the throughput. The real constraints are changeover and review of flagged parts, not acquisition.
Do I need a telecentric lens for this inspection?
Not for screening. A conventional lens is fine while wrinkles are judged by presence, size and position in the field. A telecentric lens earns its cost when you also measure — fringe width, band position or deviation in millimetres — because it removes perspective error across a 350 mm width and pairs with collimated telecentric illuminators. Start standard; upgrade when the numbers must be dimensional.