Four Doors, Two Closures: Multi-Light Fusion On Body Panels

Overview

A painted door is inspected for defects that are optically incompatible with one another. A scratch on a clear coat changes the local slope and is revealed by a directional reflection; a sanding mark under the paint is a texture at a different spatial scale; a pinhole is a dark point surrounded by gloss; a dull patch is a change in the width of the specular lobe. No single illumination shows all four at once, and any light that makes one obvious tends to suppress another.

The answer used on production lines is to image the same area under several illumination modes and combine the results. That combination is what the term multi-light fusion refers to, and it is a decision problem rather than an optics problem: which modes, how they are combined, and at what point in the pipeline the combination happens.

This article covers the modes, the two places where fusion can occur, and the calibration that decides whether a fused station holds its performance. Big Bird Industrial (formerly Shimada Big Bird Industrial) has been building cleaning and deburring systems for automotive powertrain components since 2001, and the inspection approach described here follows production stations for closures and exterior panels.

1. Why One Light Cannot See Everything

The physics is that a defect changes different optical properties, and each illumination mode is sensitive to a different one.

Slope, Texture And Absorption

A scratch is a local change in surface slope, which is best revealed by a light source at a specific angle relative to the camera. A texture defect such as orange peel is a variation in slope at a much coarser scale, which appears as a pattern in the reflected image rather than as a line. A pinhole absorbs light rather than reflecting it, which is visible against a bright background regardless of direction.

The Conflict

A source that produces strong specular contrast on a glossy panel also produces a bright band that hides anything inside it. Diffuse illumination removes the band and flattens the defect contrast. That trade is the reason a station needs more than one mode: each mode is a different compromise between contrast and coverage.

2. The Modes

Four modes cover the defect classes that appear on a painted closure, and they are normally implemented as a switchable array rather than four separate stations.

Bright Field And Dark Field

Bright field illumination places the source near the camera axis so that a smooth surface reflects into the lens. It renders gloss and colour faithfully and makes dull patches and colour deviations obvious. Dark field places the source at a low angle so that a smooth surface reflects away from the camera and only features that scatter light are bright. Scratches, sanding marks and edge defects light up against a dark background, which is why dark field is the mode that finds fine mechanical damage.

Coaxial And Grazing

Coaxial illumination sends light along the optical axis through a beam splitter, which produces an even image of a flat surface and highlights depressions as dark rings and bumps as bright centres. Grazing illumination arrives almost parallel to the surface, which reveals waviness and dents by their shadow rather than by their reflection. On a curved door skin, the grazing mode has to follow the curvature, which is why the source on such a station is often shaped to the panel rather than arranged as a flat array.

3. Where Fusion Happens

Combining four images can happen in two places, and the choice changes both the algorithm and the acceptance test.

Decision-Level Fusion

Each mode is processed independently, each produces its own candidate detections, and the results are merged by a rule. The advantage is that each mode can be tuned and validated on its own, and a failing mode can be identified without retraining the others. The disadvantage is that a defect visible only weakly in two modes may not exceed the threshold in either.

Image-Level Fusion

The modes are combined into a single representation before detection, which allows a weak signal in one mode to reinforce a weak signal in another. The advantage is sensitivity on defects that no single mode shows clearly. The disadvantage is that the combination is harder to explain, and that a change in one mode propagates into the output in a way that is not obvious when a station starts producing false calls.

Most production stations use decision-level fusion for the defect classes that are well defined and image-level fusion for the classes that are subtle, which means a station can use both approaches on the same part.

4. What Fusion Buys On A Closure

The practical benefit is coverage of defect classes that would otherwise require separate stations, and the clearest case is the combination of a highly reflective scratch with a diffuse casting or sanding defect on the same panel.

One Presentation, Several Classes

A door is presented once, and all four modes are captured in the same position. That removes the re-location error that appears when a part is inspected in two stations, and it makes the result easier to attribute because every mode sees the same surface at the same moment.

The False-Call Effect

A defect that is only weakly visible in one mode produces many false calls, because the threshold has to be set low to catch it. When a second mode confirms or rejects the candidate, the threshold can be raised without losing detection, and the false-call rate falls without touching the miss rate. That is the real benefit of fusion on a painted panel: not more sensitivity, but a better trade between the two error rates.

5. Cost And Complexity

Four modes mean four light sources, four exposures and a control system that can switch between them within the cycle. On a station inspecting several panels per minute that is a real engineering constraint, and it is why the modes are usually arranged as a ring or a shaped cowl around the camera rather than as separate optical paths.

The second cost is processing. Four images per view means four times the data and, on decision-level fusion, four detection passes. That is why the cycle time on a multi-mode station is set by compute as often as by motion, and why the point count per panel is a design decision rather than a target.

6. Calibration Across Modes

A fused station is only as good as the registration between its modes. If the dark field image is offset from the bright field image by a few pixels, a candidate found in one lands on a different location in the other, and the fusion rule either fails to confirm it or confirms the wrong thing.

Registration And Its Drift

Registration is established at commissioning with a target and drifts as the light sources age at different rates and as the mechanical assembly settles. A shift check with the same target catches the drift before it becomes a false-call problem, and it is the single most useful maintenance action on a multi-mode station.

Intensity Balance

The modes also have to stay balanced in intensity. A dark field source that dims progressively reduces the contrast of the scratches it is meant to reveal, which shows up as a gradual loss of detection on one class while the others remain normal. Recording the intensity of each mode at commissioning and checking it on a schedule turns that decline into a measurement.

7. When Not To Fuse

Fusion is not free, and on a single defect class with a stable appearance it adds cost for no benefit. A station that inspects only for missing paint or a colour mismatch does not need dark field, and a station that inspects only for scratches does not need bright field.

The decision is driven by the defect list. Where the list contains classes with incompatible optical signatures, fusion pays. Where it contains one class, a single optimised mode is cheaper to build, easier to validate and simpler to maintain.

Frequently Asked Questions

Does Multi-Light Fusion Improve Detection Rate?

It improves the trade between detection and false calls. Detection alone can be pushed high with any single mode at a cost in false calls; fusion allows both to be reasonable at once.

How Many Modes Are Needed?

As many as the defect list requires and no more. Four covers the common classes on a painted closure; a station inspecting a single class may need one.

What Is The Main Maintenance Risk?

Registration drift between modes, followed by intensity drift. Both are caught by a shift check against a reference target rather than by watching the output.

Can Fusion Be Added To An Existing Station?

Only if the mechanical arrangement allows another source at the right angle and the cycle has room for another exposure. On most single-mode stations the constraint is the geometry, not the software.

Conclusion

Multi-light fusion exists because defects on a painted panel change different optical properties, and no single illumination makes all of them visible. Combining modes buys a better balance between missed defects and false calls rather than more raw sensitivity.

The engineering effort sits in registration and intensity stability, not in the fusion rule. A station that checks both on a schedule holds its performance; one that does not drifts into false calls that nobody can explain.