Ghost Images 101: Those Reflections Aren't Paranormal

The 30-Second Version: Ghosts in machine vision are just physics. Fresnel reflections bounce around inside filter glass, lowering contrast and creating faint echoes of your image. Coatings are the exorcism.

Where Ghosts Come From

Ghosts in machine vision are usually not paranormal—they're Fresnel reflections doing physics things.

Any time light crosses a boundary between materials with different refractive indices (like air to glass, or glass to air), some light reflects instead of passing through. This is fundamental physics, and there's no way around it—only ways to minimize it.

For uncoated average glass, the rule of thumb is that about 4% of incident light reflects at each surface. The exact value depends on wavelength and angle of incidence, but 4% is the working number that optical engineers carry around in their heads.

A filter has two surfaces. That means two opportunities for reflection, and—more importantly—the potential for light to bounce between those surfaces before escaping back into your optical system.

What Ghosts Actually Do

When light bounces around inside filter glass, it can re-enter your imaging path in unintended ways. The results show up in your images as several related problems.

Reduced contrast. Even when you don't see a distinct ghost, those extra light paths elevate your black level. Areas that should be dark get a little brighter. The difference between light and dark areas shrinks. Your contrast suffers.

Milky image quality. This is the technical description of what happens when scattered and reflected light creates an overall haze across your image. Nothing looks quite as crisp as it should.

Secondary images. This is the classic "ghost"—a faint, displaced echo of a bright feature in your scene. The geometry is sneaky: a bright feature reflects off the back surface of the filter, then off the front surface, and lands on the sensor slightly offset from where it should be. You see a dim copy of your object, usually defocused because it traveled a different optical path.

Algorithm failures. This is where the machine vision consequences get real. Edge detection relies on contrast. Thresholding relies on clean separation between light and dark. Barcode reading relies on crisp transitions. OCR relies on well-defined character boundaries. When ghosts elevate your black level and wash out your edges, all of these algorithms suffer.

The Geometry Is Sneaky

Ghost images are often displaced and defocused, which makes them tricky to diagnose.

The displacement happens because the reflected light travels a different path than the direct light. It enters the filter, reflects off the back surface, travels forward through the glass, reflects off the front surface, and finally exits—but at a different angle than if it had passed straight through. On your sensor, that means the ghost appears offset from the real image.

The defocus happens because the reflected light traveled extra distance through the glass, which changes its effective optical path length. The ghost isn't in focus even when the real image is.

This combination—offset and defocused—means ghosts often appear as soft halos or displaced blurs rather than sharp duplicate images. They're easy to mistake for other problems: lens flare, sensor issues, or "the algorithm acting weird."

The Exorcism: Anti-Reflective Coatings

If surface reflections are the problem, reducing those reflections is the solution. That's what anti-reflective (AR) coatings do.

AR coatings are made from thin dielectric films—layers of material with carefully controlled thickness and refractive index. The physics is similar to what makes interference filters work: by choosing the right layer thickness, you can create destructive interference for reflected light, so that reflections from the top and bottom of the coating layer cancel each other out.

The result is dramatically reduced surface reflectance. Instead of 4% reflection per surface, a good AR coating might give you less than 0.5%. That's an 8× reduction in reflected light, which translates directly into less ghosting and higher contrast.

Coating Options and Tradeoffs

Not all AR coatings are created equal.

Single-layer coatings (like magnesium fluoride or calcium fluoride) are the entry point. They work well over a moderate wavelength range and offer a solid balance of performance and cost. If your application uses a single LED wavelength or a narrow band of wavelengths, a single-layer coating optimized for that range may be all you need.

Multi-layer coatings use multiple thin films with different refractive indices to achieve better performance over a broader wavelength range. If you're working with a wide spectral band—or if you need the best possible performance—multi-layer coatings are the upgrade. They cost more, but they perform better.

Industrial Coating Requirements

Here's something that matters in machine vision applications but might not come up in a laboratory optics discussion: durability.

Industrial filters face challenges that lab optics don't. Filters get touched during installation. They get cleaned—sometimes gently, sometimes not. They're exposed to dust, oil mist, humidity, and temperature swings. The factory floor is not a cleanroom.

When specifying filters for production, don't just ask about spectral performance. Ask about coating durability. Good industrial coatings need to be hard enough to resist abrasion from cleaning, chemically stable enough to resist whatever's in your factory air, and robust enough to survive handling by technicians who have more urgent problems than babying optical components.

A coating that delivers 0.2% reflectance in the lab but scratches when you wipe it with a lens tissue is not a good coating for production use.

Design Strategies to Minimize Ghosts

Beyond specifying AR coatings, there are several design practices that help minimize ghosting in your optical system.

Coat both surfaces. Make sure your filters have AR coatings on both sides, not just one. Each uncoated surface is a 4% reflector.

Consider coating quality in your budget. Cheaper uncoated filters might seem like a cost savings, but if they degrade your image quality, you're paying for that degradation somewhere else—in algorithm tuning, in false rejects, in customer complaints.

Minimize filter count. Every filter you add to the optical path is two more surfaces with reflection potential. If you can achieve your spectral goals with one filter instead of two, you've eliminated two ghost opportunities.

Shield against stray light. Ghosts don't just come from the direct imaging path. Bright light entering the system from unexpected angles can reflect off filter surfaces and find its way to the sensor. Proper baffling and shielding help.

Test with bright features. Ghosts are most visible when there's a bright object or specular reflection in your field of view. When testing your system for ghost problems, include some high-contrast features and look for their displaced echoes.

Quick Diagnostic Checklist

Before blaming your algorithm for poor performance, check your optics.

Are all your filters AR-coated on both surfaces? Are the coating specs appropriate for your wavelength range? Is there stray light entering the system from angles you haven't accounted for? Are there bright features in your scene that could create visible ghost echoes? Have you tested with the filter removed to confirm the filter is the source of the problem?

Sometimes the "algorithm problem" is actually a ghost problem, and the fix is a better filter—not more algorithm tuning.

The Bottom Line

Ghosts aren't mysterious. They're physics—Fresnel reflections bouncing around inside the glass elements in your optical path. Every uncoated surface is a 4% reflector. Every reflection that finds its way back to your sensor degrades contrast and creates potential ghost images.

AR coatings are the solution. They reduce reflections, improve contrast, and eliminate (or at least minimize) ghosting artifacts. When specifying filters for machine vision, coating quality should be part of the specification—not an afterthought.

And remember that industrial coatings need to be durable as well as optically good. The factory floor isn't gentle.

This post is part of KUPO's technical education series on optical filters for machine vision. Questions about filter selection for your application? Contact our optical engineering team.