cialistas.com

Taming glare and reflection in machine vision for polished lenses

Polished glass and polymer lenses behave like mirrors to a vision system. The very surface finish that makes a lens optically valuable also makes it the worst possible subject for a standard camera. Specular highlights blow out pixels, hide sub-micron scratches, and turn coating blemishes into ghosts that drift across the frame as the part moves. Every imaging engineer working on eyewear, camera optics, or laser components has watched a perfectly good lens fail inspection under a single bright LED.

Demand for reliable inspection has grown quickly across Australian precision manufacturing. Adelaide's defence optics cluster, Sydney's medical-device makers, and the eyewear producers around Melbourne all need to certify coatings, surface roughness, and edge finish before products leave the dock. With exports of medical and scientific instruments rising and the local photonics industry courting new contracts, the cost of a missed defect on a polished lens has climbed sharply. A single scratched aspheric surface can scrap a six-figure order if it slips through.

The remedy is rarely a single piece of hardware. A robust vision system for reflective parts blends carefully chosen lighting geometry, polarisation, optical filtering, sensor technology, and image-processing software. The sections that follow walk through each layer of that stack, then close with practical guidance for engineers specifying systems for Australian facilities.

Why polished surfaces defeat standard imaging

Specular reflection dominates when surface roughness drops below the wavelength of light. A finished lens at Ra 1 nm or better reflects light in a narrow cone equal to the angle of incidence, rather than scattering it back toward the camera. To a sensor, the part looks either brilliantly white or pitch black, depending on the angle, and any defect that interrupts the smooth surface shows up only as a faint, moving artefact.

Standard bright-field lighting, the default in many factories, places lamps at an angle that sends glare straight into the lens. On a matte plastic part, this reveals texture; on a polished lens, it produces a hot spot that masks half the field of view. As the part rotates on a conveyor or indexing table, the highlight slides around, creating flickering that confuses thresholding algorithms and forces engineers to slow the line.

Polished coatings and anti-reflection layers are themselves designed to manage light. Their thin-film structure interacts differently with different wavelengths, and the very feature the inspector needs to verify, such as coating uniformity, can only be seen by capturing the light that does not reflect. The default ring light or dome light simply cannot reach the information that matters.

Lighting, polarisation, and wavelength control

A diffuse dome is often the first step for general surface inspection. Light enters through a translucent hemisphere, creating a near-Lambertian bath that softens shadows and reduces highlights. On highly polished lenses, however, a dome alone tends to produce a soft, drifting glare that still hides fine defects. It works best as a baseline, not a final answer.

Coaxial lighting, where illumination travels along the same axis as the camera through a beam splitter, lights the part uniformly and eliminates silhouette shadows. It shines on planar parts such as wafers and lapped flats, but on a curved lens it produces a central hot spot that mirrors the lens curvature. Engineers often pair it with a slightly off-axis camera or a tilted beam splitter to break the perfect symmetry.

Dark-field illumination, with light coming in at a steep angle from the sides, makes scratches, edge chips, and dust flare brightly against a dark background. It excels at revealing micro-defects on otherwise mirror-like surfaces, which is exactly what polished lens inspection needs. Many production lines run dark-field as the primary inspection mode, using a second bright-field channel for gross-feature verification.

Cross-polarised lighting adds another lever. By placing a polarising filter in front of the light source and a crossed analyser in front of the camera, specular glare can be suppressed by more than 90 percent, while diffuse light from scratches and contamination still reaches the sensor. A pair of crossed linear polarisers acts as a tunable glare dial: loosen the analyser a few degrees and the defect contrast shifts in a predictable way. Circular polarisers handle parts that rotate under the camera, where the polarisation angle would otherwise drift.

Bandpass and longpass filters narrow the wavelength range reaching the sensor. Blue light at 450 nm resolves sub-micron features because its wavelength is short, but it also brightens surface haze. Near-infrared at 850 nm or longer can pass through some anti-reflection coatings while being absorbed by contamination, making oil and fingerprint residue stand out as dark spots. Multi-spectral imaging takes this further, capturing the part in several narrow bands and stacking the results into a single, defect-rich image. The benefits of multi-spectral inspection are most visible on coated optics, where one wavelength highlights the coating layer and another penetrates to the substrate. A combined image suppresses glare that no single channel could remove, and it flags contamination that looks identical to the substrate in white light.

Sensors, optics, and filters for reflective parts

A sensor's dynamic range matters more than its pixel count when glare is in play. Polished parts routinely deliver scene brightness spanning 1000:1 or more, and a standard 8-bit camera clips the bright end of that scale. Modern global-shutter CMOS sensors with 12-bit or 14-bit ADCs, and high-dynamic-range modes that combine short and long exposures, preserve the defect contrast in the dark regions while keeping the highlight from saturating. A careful CCD and CMOS comparison for this kind of work shows that the choice now depends on read noise, quantum efficiency at the chosen wavelength, and shutter behaviour, rather than on raw sensitivity alone.

Telecentric lenses are the standard for any metrology of polished parts. They remove perspective distortion so a scratch on a curved surface appears the same size whether it sits at the centre or the edge of the field. They also produce a uniform magnification across the working distance, which matters when the lens is not sitting perfectly flat on a stage. For very small features, a long working distance and a high-resolution lens cost depth of field, and a beam splitter with coaxial lighting can give the inspector a stable view.

Filters deserve the same selection rigour as the lens itself. A polariser should be a quality coated glass type rather than a polymer film, because film polarisers shift transmission with angle and burn out under high-intensity strobes. Neutral density filters balance bright and dark regions in HDR captures. Mounting all of these in a single, vibration-isolated optical bench keeps the setup aligned when the line runs at 200 parts per minute.

Software, system design, and Australian field realities

Software can recover a great deal from images that look hopeless on screen. High-dynamic-range compositing merges a short exposure of the highlight with a long exposure of the defect. Polarimetric imaging captures several frames at different analyser angles and computes a per-pixel degree of polarisation, which separates specular from diffuse components almost perfectly. Many Australian integrators now ship these as off-the-shelf libraries, because the alternative, custom image processing per line, is too slow to maintain.

When designing a vision system for Australian plants, engineers must account for ambient light rejection, sealed enclosures rated to IP65 or better, and thermal management for the LED arrays that otherwise drift in colour temperature as the line warms up. Sydney's harbour-side factories deal with high humidity and salt spray, Brisbane sites in tropical Queensland see intense summer sun streaming through roller doors that operators leave open for ventilation, and Perth's mining-optics suppliers ship parts that have crossed the Nullarbor on dusty roads. Adelaide's defence optics workshops run climate-controlled rooms that still struggle with the local 40°C summer peaks.

Local regulation also shapes the choice. The AS/NZS 3820 essential safety standard for electrical equipment, AS/NZS IEC 60825 for laser products, and the Work Health and Safety Act 2011 all influence how lighting enclosures and laser-assisted inspection are specified. For plants exporting under a CE or FDA regime, the same system often has to satisfy three sets of rules at once. Building for the strictest standard up front saves money later.

Lighting technique Glare suppression Best defect type revealed Strengths Limitations
Diffuse dome Moderate General surface haze, contamination Simple, forgiving of part tilt Soft glare on mirror surfaces
Coaxial bright-field Low to moderate Edge geometry, planar features No silhouette shadows Central hot spot on curved lenses
Dark-field High Scratches, edge chips, dust Reveals micro-defects Hides form and roughness
Cross-polarised Very high Coatings, subsurface features Tunable, works at any angle Reduces overall light, needs careful alignment
Multi-spectral High Coating defects, mixed-material flaws Works on hard cases Higher cost, more data to process

Practical guidance for specifiers

For engineers ready to specify or upgrade a system, a few principles consistently separate a working setup from a frustrated one. The order of choices matters: lighting geometry sets the ceiling for what the rest of the stack can achieve.

The most common mistake is to start with the camera. Even an expensive 20-megapixel sensor will produce unworkable images if a bright LED is pointed at a polished part. Investing in a well-engineered lighting enclosure, vibration-isolated mounts, and quality filters pays back faster than any sensor upgrade, and a measured validation run on real production parts will reveal weaknesses no datasheet can predict.