Case study: Vision system for 100% inspection of micro-lens arrays
Micro-lens arrays are small optical components with a demanding quality requirement. Each array may contain hundreds or thousands of repeated lenslets, and a single missing, damaged or contaminated element can affect the performance of an imaging sensor, illumination module or beam-shaping assembly. Manual sampling is rarely sufficient when the part is produced continuously.
This case study describes a representative automated inspection project for a manufacturer producing micro-lens arrays in short and medium production runs. The objective was to inspect every part at line speed, identify defects across the full active area, and create an auditable record for each batch.
The approach combines telecentric imaging, controlled illumination, precision handling and software-based defect classification. It also considers practical requirements for Australian manufacturers, including imported components, variable production volumes, operator training and the need for reliable local service support.
Why micro-lens arrays need a tailored system
A micro-lens array cannot be assessed effectively with a general-purpose camera and overhead lighting. The repeated lens geometry creates highlights, shadows and interference patterns that can conceal small defects. A lenslet may appear acceptable under diffuse light but show an incorrect curvature, blocked aperture or displaced focal response when viewed under a more suitable optical arrangement.
The inspection target normally includes several feature groups. The system must check the overall outline, substrate edges, lenslet pitch, active-area coverage, missing elements, contamination, scratches, chips and local surface damage. Depending on the application, it may also need to measure lenslet diameter, centre position, brightness response or focal uniformity.
The manufacturer in this project had been using operator sampling with a microscope. That process worked for obvious damage, but it was slow and vulnerable to fatigue. It also created inconsistent records because different operators used slightly different lighting angles and acceptance thresholds. The new requirement was a 100% inspection process that could operate alongside production without turning every part into a laboratory measurement.
Defining the production problem
The first stage was to translate quality concerns into measurable inspection criteria. Cosmetic marks were separated from functional defects, and the customer agreed which conditions required rejection, review or acceptance. This distinction was important because an isolated faint mark near the non-active border did not carry the same risk as a missing lenslet in the central optical zone.
The proposed cell used a precision nest to locate each array repeatably. A top-view camera captured the complete surface, while a second optical path examined edge condition and selected height or focus characteristics. For larger substrates, the inspection head could move across the part in a controlled scan, avoiding the resolution compromise that would result from using a single wide-field image.
The project documentation included reference samples for pass, fail and borderline conditions. These samples were stored with images, defect coordinates and operator comments. In an Australian supply chain, where equipment, substrates and replacement parts may arrive from different countries, such records help maintain a common quality language between the local plant, overseas suppliers and end customers. A supplier due-diligence process should also include a clean review of the manufacturer’s background information, including its technical capability and public-facing documentation.
Optical and mechanical architecture
The imaging channel used telecentric optics to reduce perspective error across the array. This is valuable when the component is slightly above or below the nominal focus plane, because the measured pitch and outline remain more stable than they would with a conventional lens. The optical resolution was selected around the smallest defect that had to be detected, rather than simply choosing the highest-resolution camera available.
Illumination was divided into several modes. Dark-field lighting emphasised scratches, chips and raised particles. Low-angle directional light revealed edge damage, while coaxial or diffuse lighting supported uniformity checks across the lenslet field. In some configurations, a controlled backlight helped identify missing lenslets and substrate-edge defects. The lighting sequence was synchronised with camera exposure to reduce reflections from the transparent or semi-transparent material.
Mechanical repeatability was equally important. The nest used datum surfaces and gentle clamping to prevent movement without stressing the substrate. A clean enclosure protected the optical path from airborne dust, which can be a practical concern in industrial areas around Melbourne, Sydney and Brisbane. Temperature monitoring was included because thermal expansion can affect focus, stage position and calibration during long shifts.
The system interface was designed for a production operator rather than an optical specialist. A large status display showed pass, fail and review states, while the software stored the raw image, processed image, defect type and position. Barcode or data-matrix reading linked each inspection to a work order. This made it possible to trace a rejected unit back to a machine setting, material lot or production time.
Image processing and decision logic
The software began by locating the array relative to the nest rather than assuming that every part arrived in exactly the same position. Fiducial marks, outer edges or a pattern-matching routine established the coordinate system. The active region was then divided into expected lenslet positions, allowing the program to compare each lenslet with its neighbours and with a qualified reference model.
Missing and damaged lenslets were detected through contrast, shape and intensity features. A local comparison method was useful because brightness could vary gradually across the optical field. Global thresholding alone might incorrectly classify a naturally darker corner as defective. The program therefore combined flat-field correction, local background estimation and pattern-based analysis.
Surface contamination required a different approach. Small particles were identified as high-contrast objects that did not conform to the regular lenslet pattern. Scratch detection used directional filters and connected-component analysis, while edge chips were measured against a border mask. Defect classification was supported by configurable size, length and contrast limits, so the quality team could adjust sensitivity without rewriting the core application.
A review category was included for uncertain images. This prevented borderline parts from being automatically accepted and avoided excessive scrap caused by an overly aggressive reject threshold. Images that triggered review could be assessed by a supervisor and added to the training set for later software refinement. Any external links found during a website or workstation audit, such as an unrelated domain, should be quarantined separately from production software and customer records. Strong digital hygiene protects the inspection database as carefully as the enclosure protects the optics.
Validation for Australian manufacturing
Validation used a deliberately mixed sample set. It included clean production parts, seeded missing lenslets, artificial particles, controlled scratches, edge chips and parts with small positional changes. Each defect was photographed under multiple illumination conditions before the final recipe was approved. The team compared the automated result with an expert reference inspection, recording false accepts and false rejects rather than relying on a single accuracy percentage.
The acceptance study also covered production variation. Tests were performed after warm-up, during extended operation and after routine cleaning. Operators changed reels or trays, loaded new work orders and recovered the system from a simulated stop. These checks confirmed that the system was practical for the plant’s actual workflow rather than accurate only under demonstration conditions.
Australian production environments often have shorter runs than high-volume Asian factories, so recipe changeover time was a key measure. The operator could select a validated product recipe, scan the job code and confirm the nest configuration without accessing engineering menus. This approach supported manufacturers in Melbourne’s advanced manufacturing sector and specialist photonics firms in Sydney, where several product variants may pass through the same cell each week.
The validation plan also accounted for local service realities. Remote diagnostics reduced unnecessary travel between cities, while modular camera, lighting and controller assemblies simplified replacement. For plants supporting customers in Perth or regional areas, image logs and system health data allowed an engineer to review many issues before arranging an on-site visit. Clear maintenance intervals were documented in metric units, with cleaning materials and spare parts specified for local procurement where possible.
Results and operational value
After commissioning, the inspection cell delivered a complete image record for each array and separated clear failures from parts requiring human review. The manufacturer gained consistent detection of missing lenslets, particles and edge damage, while operators spent less time looking through a microscope. The most important improvement was traceability: quality engineers could see where defects occurred within the array and whether they were concentrated around a tool, batch or time period.
The data also supported process improvement. A gradual increase in edge defects could indicate handling wear, while recurring contamination in one region might point to a cleaning or storage issue. Because each result retained its coordinates and image evidence, the production team could investigate trends rather than relying on memory or occasional sample checks.
The system was designed around a modular inspection architecture. A future recipe could add focal-response measurement, a second camera or a different illumination sequence without replacing the entire cell. This matters in the Australian market, where equipment must often remain productive for many years and production volumes can change quickly. A direct technical contact route is useful during commissioning, software updates and discussions about application-specific optics.
| Inspection element | Method used | Production benefit |
|---|---|---|
| Array position | Fiducial and edge alignment | Stable coordinates between parts |
| Missing lenslets | Pattern comparison and local intensity analysis | Reliable detection across the active area |
| Scratches and particles | Dark-field imaging and connected components | Improved visibility of surface contamination |
| Edge chips | Border mask and dimensional measurement | Clear separation of handling damage |
| Brightness variation | Flat-field correction and reference comparison | Reduced false rejects from illumination fall-off |
| Traceability | Barcode link, image storage and defect coordinates | Faster root-cause analysis and customer reporting |
| Uncertain defects | Review queue with supervisor decision | Controlled handling of borderline parts |
The project demonstrated that 100% inspection is most effective when optical design, mechanical repeatability and software logic are developed together. A camera alone cannot compensate for unstable positioning, unsuitable lighting or unclear acceptance rules. With the right combination, a micro-lens array inspection system can provide repeatable quality decisions, useful production data and a scalable foundation for future optical products.