Automated Scratch and Pit Detection for Large Optical Windows
Large optical windows are difficult to inspect by eye because their size, curvature, coatings and reflective surfaces can hide small defects. A scratch that is obvious under one angle may disappear under another, while a shallow pit can resemble dust, coating variation or a highlight from the illumination system. Automated machine vision brings repeatable lighting, imaging and measurement to this quality-control task.
A well-designed inspection cell can locate, classify and record surface imperfections across glass panels, protective windows, infrared covers and other precision optics. For Australian manufacturers and distributors, the system also needs to suit local production volumes, long supply chains, workplace practices and the practical realities of servicing equipment in Sydney, Melbourne, Brisbane or regional sites.
| Inspection approach | Strengths | Typical limitations |
|---|---|---|
| Manual visual inspection | Flexible and inexpensive to start | Inconsistent results, operator fatigue and weak traceability |
| Bright-field machine vision | Effective for many visible scratches and contamination marks | Highly reflective surfaces can produce glare |
| Dark-field inspection | Makes fine scratches, pits and edges stand out | Requires careful control of stray light |
| Multi-angle automated inspection | Improves detection across changing defect orientations | Higher cost and more complex calibration |
| 3D or profilometric inspection | Measures depth, height and surface profile | Slower and often more demanding to integrate |
Why Large Windows Need Specialized Inspection
Large-format optics create a field-of-view problem. A camera with insufficient resolution may cover the complete window but fail to resolve a fine linear mark. A high-resolution camera may reveal tiny defects yet require several overlapping images, precise motion control and software that can stitch or compare those images accurately.
The defect itself also affects the inspection method. Scratches are generally elongated and directional, while pits are localised depressions or chips that scatter light in a different pattern. Edge damage, coating pinholes, haze and embedded particles add further classes that should be separated rather than placed into one broad “reject” category.
Human inspection remains useful for investigating unusual results, but it is a poor sole standard for continuous production. Operators can become accustomed to minor marks, interpret acceptance limits differently and struggle with large panes under changing ambient light. Automated imaging creates a stable record tied to a part number, batch, position and inspection result.
Choosing Illumination and Optics
Illumination is usually more important than camera resolution. Bright-field lighting sends reflected or transmitted light directly toward the camera and can reveal broad contamination, coating non-uniformity and contrast changes. Dark-field lighting places the main reflection outside the camera’s view, allowing scratches, pits and small chips to appear as bright scattering features against a darker background.
Large windows often need several lighting geometries. A low-angle bar light can emphasise linear scratches, while diffuse dome lighting reduces harsh reflections from a curved or coated surface. Polarising filters may suppress unwanted glare, although they can also reduce the signal from certain defects. Changing the incident angle between image captures is valuable when a mark is visible only along a particular direction.
The lens must be selected around working distance, sensor size and required ground sampling distance. Telecentric optics can reduce perspective error and make dimensional measurements more stable, but they may be physically large and expensive for oversized windows. In other installations, several calibrated cameras provide better coverage than one extreme-resolution camera. A supplier’s broader optical inspection range can help when comparing lenses, sensors and inspection modules for this trade-off.
Building the Inspection Cell
A robust cell starts with mechanical stability. The window should be supported without creating stress marks or allowing vibration during image capture. Vacuum fixtures, soft contact points and adjustable frames can be used according to the material and shape. The support design must also leave enough access for lighting to reach the inspected surface from useful angles.
Motion can be continuous or step-based. A conveyor or gantry is suitable when windows arrive in a predictable orientation and the system can synchronise encoder position with image capture. Step-and-shoot operation is often easier for large panels because it allows the stage to settle before each exposure. Overlapping fields of view are essential; otherwise, defects may be lost in the narrow boundaries between images.
The inspection area should be enclosed from direct sunlight and uncontrolled factory lighting. This matters in Australia, where a western-facing roller door or strong summer daylight can change the appearance of reflective glass during a shift. A darkened enclosure, controlled backgrounds and interlocked access doors improve repeatability and protect operators from moving equipment.
Turning Images into Reliable Defects
Image processing normally begins with flat-field correction, lens distortion correction and background normalisation. These steps reduce variation caused by sensor pixels, illumination fall-off and small differences between camera channels. The software can then segment bright or dark features, connect nearby pixels and calculate length, width, area, orientation and contrast.
Scratch detection benefits from directional filters and line-enhancement methods. A system may inspect the same region with filters aligned horizontally, vertically and diagonally, then combine the results. Pits can be found through blob analysis, local contrast, shape measurements or a height map where depth information is required. Thresholds should be based on validated samples rather than an arbitrary grey-level value.
Machine learning is useful when defects vary widely or when acceptable cosmetic variation is difficult to describe with fixed rules. A model needs images of genuine scratches, pits, dust, bubbles, coating marks and clean surfaces. Practical guidance on machine-vision training is relevant here, especially for building a labelled image set and separating training data from independent validation data.
The decision should remain explainable to production staff. A result showing the defect location, category, confidence and cropped image is easier to review than a simple red or green signal. For regulated or high-value optics, the original image and measurement record should be retained with a timestamp and serial number.
Managing Scale, Speed, and Data
An oversized window can generate hundreds of high-resolution images per inspection. The vision controller therefore needs sufficient processing power, fast storage and a network architecture that does not lose images while a line is running. Local processing can deliver an immediate pass or fail, while selected images and statistics are sent to a manufacturing execution system or quality database.
Speed should be calculated from the full cycle, rather than camera exposure alone. Loading, alignment, settling, image capture, stage movement, analysis and operator review all contribute to takt time. A slower scan with high confidence may be cheaper than a fast system that creates frequent false rejects and requires manual rechecking.
Defect maps are especially valuable for large optical windows. They show whether marks cluster near a fixture, occur along a coating path or increase toward one side of a sheet. This information can identify problems in handling, cleaning, polishing, lamination or transport. A supplier that presents case studies and measurable outcomes deserves closer attention than one that lists only camera specifications.
Data retention also needs a practical policy. Keeping every full-resolution frame forever may be unnecessary, while deleting evidence immediately can make disputes impossible to investigate. Many factories retain all rejected images, a sample of accepted parts and summary statistics for each batch, with longer retention for safety-critical products.
Validating Performance in Australian Production
Acceptance criteria should be written before equipment is purchased. Define the smallest scratch, pit or chip that must be detected; specify its location, length, diameter, depth or contrast; and state which cosmetic variations are acceptable. Include edge zones and coated areas, since these often behave differently from the centre of the window.
Testing should use a representative defect library. Genuine production samples are preferable, supplemented by controlled reference specimens with known marks. The validation protocol should measure probability of detection, false-reject rate, repeatability and reproducibility between shifts. Repeating the test after cleaning the optics and restarting the system helps expose hidden dependencies.
Australian operations may need to coordinate equipment delivery across long distances, from a Melbourne plant to a remote service team or from a Brisbane integrator to a customer in Perth. Spare cameras, light modules and cables should be identified early, with clear arrangements for remote diagnostics and replacement parts. Local electrical, guarding and workplace safety requirements must be addressed during installation rather than after commissioning.
Procurement teams often compare capital cost with labour savings, but quality risk has equal importance. A missed defect can lead to field returns, rework and reputational damage, while an excessive false-reject rate consumes good material. A measured business case should include inspection throughput, operator hours, scrap, warranty exposure and the value of searchable process data.
Integrating Operators and Quality Teams
Automation works best when operators can understand and challenge its decisions. A large display can show the inspected region, defect category and acceptance threshold. Controls should support secure override procedures, with every override logged against a user and reason. This prevents informal threshold changes from becoming invisible process drift.
Training should cover cleaning, fixture loading, basic fault recovery and the meaning of defect classes. A vision system may be technically accurate yet perform poorly if fingerprints remain on the support surface or if the window is loaded several millimetres off position. Simple work instructions, photographs and routine checks are often more useful than a long technical manual.
The system should distinguish inspection failure from product failure. A dirty lens, blocked light, missing image tile or communication error needs a maintenance alarm, not an automatic assumption that the optical window is defective. Health checks can use reference targets, monitored light output and image-quality statistics to identify gradual degradation.
For businesses assessing adjacent vision applications, an optical systems reference can be useful when mapping requirements across sensors, lighting and measurement tasks. The important procurement question is whether the proposed platform can be adapted without replacing the entire cell when a new coating, product size or defect class is introduced.
Maintaining Accuracy Over Time
Calibration should be scheduled around actual risk. Camera alignment, lens focus, stage position and light intensity can drift through vibration, temperature changes and routine cleaning. A reference artefact with stable features provides a quick daily or shift-start check, while a more complete calibration verifies scale, distortion and camera-to-camera registration.
Environmental control matters for glass and coated optics. Dust, airborne oil and cleaning residue can create marks that look like pits or scratches. The inspection cell should have a defined cleaning method, suitable gloves and lint-free materials, with records showing when cameras, windows and fixtures were serviced. In coastal areas such as Sydney or Brisbane, humidity and airborne contaminants may require extra attention to enclosure condition and metalwork.
Software updates should be controlled like any other production change. Save the previous configuration, thresholds, model version and calibration files before applying an update. Re-run the agreed reference set afterward and record the result. If a machine-learning model is retrained, its training images, validation metrics and release approval should remain traceable.
A properly engineered system provides more than an automated reject signal. It gives manufacturers consistent surface inspection, defect location, process feedback and evidence for customer discussions. For large optical windows, the combination of controlled illumination, calibrated imaging, measured acceptance criteria and disciplined maintenance determines whether automation becomes a dependable quality tool or an expensive source of uncertainty.