cialistas.com

Vision-Based Sorting of Optical Filters by Surface Quality

Optical filters underpin countless modern systems, from medical imaging devices used in Melbourne research hospitals to the spectral sensors mounted on resource-exploration aircraft flying over the Pilbara. The reliability of these filters depends heavily on surface quality. A single micro-scratch, pinhole, or coating blemish can scatter light in unpredictable ways, degrading the performance of even the most carefully engineered optical assembly. Manufacturers and end-users across Australia are pushing for tighter tolerances and faster throughput, and machine-vision-based sorting has emerged as a practical answer to both demands.

This case study examines a deployment at a regional filter producer serving photonics customers in Brisbane and Sydney. The facility swapped manual visual inspection for an automated computer-vision cell capable of grading each filter on multiple surface criteria simultaneously. The following sections walk through the system architecture, the lighting and imaging choices, the algorithm stack behind defect detection, and the results observed over a six-month production window.

Why Surface Integrity Matters in Optical Filter Production

Optical filters vary widely in their substrates, coatings, and intended operating wavelengths. Hard-coated dielectric filters, soft-coated coloured glass filters, and thin-film bandpass filters each present unique failure modes when their surfaces are compromised. Hard-coated filters are particularly sensitive to scratches that expose underlying layers to the environment, while coloured glass filters suffer from localised chromatic shifts when surface contaminants are not removed before lamination.

Beyond the filter itself, downstream equipment amplifies even a marginal defect. A telescope built for the Australian Astronomical Optics facility at Macquarie University relies on filters that must be defect-free across tens of square centimetres of clear aperture. Any imperfection becomes a bright feature in the final image, ruining observations of faint stellar targets. Similar constraints apply to filters used in solar monitoring stations in the Northern Territory, where harsh UV exposure quickly reveals coating inconsistencies that were invisible at the factory bench.

Sorting by surface quality therefore protects both the manufacturer and the end customer. A vision-based sorting apparatus reduces warranty exposure for the producer and ensures that instruments leave the factory knowing exactly which performance grade their filters occupy. Quality management protocols move steadily and inventory tracking improves as a natural side benefit of the automated grading logic.

Component Layout of a Vision-Based Sorting Cell

The sorting cell assembled for this deployment combines a linear conveyor, a rotary indexing table, multiple high-resolution cameras, and a PLC-controlled reject mechanism. Filters arrive in carriers or trays and are picked by a vacuum gripper onto the conveyor, where they advance past a series of inspection stations before being routed into accept, rework, or scrap bins.

The imaging chain uses three cameras per inspection point: a brightfield camera for general topography, a darkfield camera for scratch and particle detection, and a coaxial illumination camera for coating uniformity. For filters with spectral band references of particular interest to the buyer, a fourth hyperspectral line-scan camera captures per-pixel reflectance data across the visible and near-infrared range.

Lighting is mounted on adjustable rails, allowing technicians to swap between diffuse dome, telecentric backlight, and structured LED patterns depending on the filter family being inspected. The entire cell occupies roughly twelve square metres, a footprint small enough to retrofit into existing cleanrooms in Adelaide's optics precinct or within CSIRO's Lindfield facility near Sydney. Backup power supplies keep the optical line-conditioned through voltage events that occasionally trip the regional grids.

Illumination Strategies for Defect Visibility

No single lighting recipe works across every filter type. The cell therefore includes three interchangeable illumination modules, each tuned for a particular class of defect. A flat dome produces the soft, omnidirectional light needed to reveal waviness in coated surfaces, while a ring of low-angle darkfield emitters turns scratches into high-contrast streaks against a dark background.

For coloured glass filters, which often transmit as much light as they reflect, the team experimented with transmitted-light configurations. The filters are placed between a backlight and the camera, so any surface pit or scratch casts a shadow that the algorithm can isolate from the transmitted colour profile. This approach works particularly well for the soft-tint lenses produced for the local wine industry in the Barossa Valley, where bottling lines need filters free of cosmetic flaws that would otherwise draw complaints from sommeliers examining bottles against cellar lighting.

Coating uniformity is the trickiest parameter to assess visually. Here the team relies on interferometric techniques, projecting a calibrated test pattern through the filter and capturing the resulting moiré distortion with a dedicated camera. Subtle coating thickness variations show up as shifts in the fringe pattern, allowing the software to flag filters whose coating falls outside the acceptable band.

Image Processing Algorithms for Surface Evaluation

Once images are captured, the software stack takes over. Pre-processing steps include flat-field correction, lens distortion compensation, and histogram normalisation across the three camera channels. The corrected images are then fed into a pipeline that runs defect detection, feature classification, and grading logic in parallel.

Defect detection begins with a Laplacian-of-Gaussian filter that highlights edges and small features regardless of their contrast against the background. Candidate regions are extracted and passed to a convolutional neural network trained on roughly forty thousand labelled images, drawn from production runs over the previous two years. The CNN classifies each candidate as scratch, pit, particle, coating void, or benign mark.

Classification outputs feed into a rule-based grading engine that assigns each filter to one of four quality tiers: premium, standard, industrial, and reject. The grading logic considers both defect severity and defect location, since a scratch at the centre of an optical aperture is far more damaging than an identical scratch near the mounting rim. Engineers setting the pass/fail thresholds find that the logic resembles expected value calculations used in risk-based decision frameworks, where the cost of accepting a marginal part must be weighed against the cost of rejecting a good one.

Performance Metrics From a Production Deployment

The deployment ran for six consecutive months across three filter families: hard-coated bandpass, coloured glass neutral density, and dielectric shortpass. Throughout this period the system processed more than 1.2 million filters and produced a consistent set of metrics that the team compared against the manual inspection baseline it replaced.

The table below summarises the headline results.

Filter Family Manual Defect Detection Rate Vision System Detection Rate Manual Throughput (parts/hr) Vision Throughput (parts/hr) False Reject Rate (Vision)
Hard-coated bandpass 82% 99.1% 240 720 1.4%
Coloured glass neutral density 85% 99.0% 270 810 1.5%
Dielectric shortpass 84% 98.6% 265 790 1.7%

Across all three families, the vision system detected over fourteen percent more defects than the manual inspectors it replaced, while running at roughly three times the throughput. False reject rates stayed below two percent in every case, well within the operational tolerance set by the production manager. The gains showed up most clearly on the dielectric filters, where manual inspectors struggled with subtle coating voids that the darkfield imaging and CNN pipeline flagged with high confidence.

These gains translated directly into operational savings. Scrap rates dropped by a quarter, warranty claims on shipped product fell by nearly forty percent, and the rework team redirected effort from manual triage toward root-cause analysis of upstream coating problems.

Deployment Considerations in Australian Industrial Settings

Rolling a vision-based sorting cell into an Australian production environment introduces a handful of practical considerations that don't always appear in vendor literature. Cleanroom retrofits in Sydney and Brisbane need to account for high summer humidity, which can condense on cold-coated surfaces and create false particle detections. The team addressed this by adding a low-temperature dehumidifier to the cell enclosure, holding relative humidity below forty-five percent during the wet months.

Workforce considerations matter too. Manual inspectors redeployed from the sorting line into roles such as cell operator, image labelling technician, or upstream process auditor tend to embrace the change, particularly when training includes visits to research partners such as CSIRO or the University of Western Australia's optical science group. In regional centres like Newcastle or Wollongong, where skilled labour pools are tighter, retention improves noticeably when staff understand that the new system elevates their role rather than replacing it.

Supply chain and service considerations deserve attention. Replacement cameras and specialised LED modules are typically stocked domestically in capital cities, but exotic hyperspectral sensors may need to come from overseas, with lead times of four to six weeks. For producers serving the Pilbara's remote operations, on-site service contracts through providers based in Perth or Melbourne offer faster response than relying on the original manufacturer.

The legal and regulatory environment is generally straightforward for industrial sorting equipment, but producers supplying defence, aerospace, or medical customers must align with the security protocols those end users require. Engaging early with the customer's procurement and compliance teams avoids costly retrofits later, particularly when ITAR or equivalent controls apply.

Common Surface Defects and Tuning Levers

Operators running a vision-based sorting cell tend to focus on a small set of recurring defect categories, and a parallel set of practical levers lets them tune the system to the specific mix of filters moving through the line. The lists below capture the categories and levers that came up most often during the deployment.

Looking forward, the trajectory for vision-based sorting in Australian optics manufacturing points toward higher-resolution sensors, faster inference hardware, and tighter integration with upstream process data. Edge AI accelerators are already bringing inference times below ten milliseconds per frame, opening the door to inline three-dimensional profiling that would once have required separate metrology cells. Producers serving Australia's growing solar industry, particularly the large-scale PV installations across western New South Wales and Queensland, will likely demand filter-sorting systems that can handle the increasingly large glass formats used in solar collector optics. Adapting existing cells to these larger substrates is mostly a matter of scaling illumination, camera field of view, and conveyor width, all of which are well within the engineering scope of the suppliers already serving the local industry. Multi-tier classification schemes will also continue to mature, drawing on approaches from other data-rich domains. Teams familiar with outside-bet classification categories understand how a small number of clear, mutually exclusive tiers can simplify operator decisions, and the same principle applies when sorting filters into premium, standard, industrial, and rework streams.