Machine vision systems explained

Lighting first, software second

Most vision projects that struggle are lighting problems wearing a software costume. If the feature you need to measure is not visibly different from its background in the raw image, no algorithm will reliably find it. Bright-field, dark-field, backlight, dome and polarised arrangements each make a different kind of feature stand out, and choosing between them is the first real decision on any inspection. Films, foils, PET, glass and metal are all readable with the right optical setup, and all unreliable without it.

Rule-based or AI

Rule-based tools measure things you can define: a distance, an edge, a colour, a presence check. They are fast, repeatable and easy to justify. AI and edge learning tools handle the cases you cannot write a rule for, like a scratch that looks slightly different every time, or characters printed on a varying background. The In-Sight 2800 combines edge learning with rule-based vision for error-proofing, and the D900 runs deep learning on the device for OCR and defect detection with no PC required.

Picking the right camera

Resolution follows the smallest feature you need to see across the field of view you need to cover, not the other way round. The In-Sight range runs from the 2000 series sensor at up to 1.2 MP and SnAPP at up to 1.6 MP, through the 2800 at up to 2 MP, the 7000 and 8000 at up to 5 MP, the 3800 at up to 16 MP, and the 9000 at up to 12 MP area scan or 32 MP line scan for measurement work. Most carry integrated lighting and autofocus and are rated IP67.

3D inspection

Some checks need height rather than appearance: fill level, seating, warp, or a bead of adhesive that is present but too thin. The L38 works at 1920 points with edge learning, the 3D-A5000 captures around 1.5 million points as an area scan, and the 3D-L4000 laser displacement sensor runs with VisionPro for custom applications. All are factory calibrated, which removes a long and error-prone step at commissioning.

Guiding robots

Vision-guided pick and place lets a robot find the part rather than requiring the part to arrive in exactly the same place every time. That removes fixtures, shortens changeovers and makes bin picking possible at all. The same offsets correct tool position on automated machine tending cells.

Codes and vision are different jobs

Vision systems can read 1D and 2D codes, but if the requirement is a printed-quality grade against a standard rather than a successful decode, that is barcode verification and scanning. It is worth being explicit about which one a customer or auditor is actually asking for, because the equipment and the evidence differ.

False rejects and stability

A system that rejects good product is worse than no system, because operators learn to override it. Fixturing, stable lighting, golden reference samples and sensible statistical thresholds are what keep false rejects down, with adaptive tolerances and a review step where the call is genuinely marginal. Keeping lenses and lighting clean, checking focus and exposure and reviewing trend data are the ongoing tasks.

Specification

Vision systems: Cognex In-Sight 2800, 3800, 7000, 8000, 9000 and D900, standalone with processing on the device
Resolutions: Up to 2 MP on the 2800, 5 MP on the 7000, 8000 and D900, 16 MP on the 3800, and 12 MP area or 32 MP line scan on the 9000
Vision sensors: In-Sight SnAPP pre-trained AI at up to 1.6 MP and In-Sight 2000 rule based at up to 1.2 MP, both with PoE
3D vision: In-Sight L38 at 1,920 points, 3D-A5000 area scan at 1.5 million points, and 3D-L4000 laser displacement, all factory calibrated
AI tools: Edge learning on the 2800 and 3800; deep learning and ViDi on the D900 for OCR and defect detection

Protection: IP67 across most In-Sight systems, IP65 on the 2000 series and 3D units; stainless housings for washdown
Lighting: Integrated lighting and autofocus; bright-field, dark-field, backlight, dome and polarised chosen for contrast
Software: In-Sight Vision Suite with EasyBuilder and spreadsheet environments; VisionPro for PC based applications
Integration: Digital I/O or fieldbus including PROFINET and EtherNet/IP for pass and fail, triggers and robot offsets
Validation and records: Image and result retention with timestamps, user IDs and lot data; IQ, OQ and PQ support and change control

FAQs about Machine Vision

Quick answers on capability, lighting, speed, validation, integration and maintenance.

What defects can vision detect?

From missing components and skewed labels to scratches, dents, voids, colour shifts and barcode/print issues—configured to your CTQs.

How fast can inspections run?

Milliseconds per part with the right optics/lighting; multi-camera cells handle high-speed packaging and continuous motion lines.

Do we need special lighting?

Usually. We specify bright-field, dark-field, backlight, dome or polarised lighting to maximise contrast on your materials.

Can vision read low-contrast or curved codes?

Yes—proper lighting, optical setup and algorithms enable reliable reads on films, foils, PET, glass and metal.

How does it integrate with our PLC/robots?

Digital I/O or fieldbus (Profinet/EtherNet-IP/etc.) for pass/fail, triggers and robot offsets; recipes select per-SKU parameters at the HMI.

Can we store images for audits?

Yes—configurable image/result retention with timestamps, user IDs and lot data for CAPA, GMP and customer reporting.

What about validation for pharma/medical?

We provide documentation, IQ/OQ/PQ support and change-control procedures aligned to your quality system.

Is vision suitable for washdown areas?

IP-rated cameras, stainless housings and sealed lighting options are available for hygienic environments.

How are false rejects minimised?

Proper fixturing, lighting stability, golden samples and statistical thresholding; we can add adaptive tolerances and AOI review where needed.

What maintenance is required?

Keep lenses/lighting clean, verify focus/exposure, review trend data and update golden references; remote diagnostics are available.

Machine Vision Systems