Manufacturing visual inspection
Computer vision for manufacturing quality inspection
Assess whether camera-based inspection can support a defined quality decision on your line. NeoBram starts with defect definitions, representative parts, imaging constraints and the process for reviewing uncertain results.
Define the defect and the consequence
Choose one station, part family and inspection purpose: surface-defect review, component presence, assembly orientation or a defined visual check. Quality engineers specify acceptable variation and the cost of a missed defect versus an unnecessary reject. Avoid combining unrelated defect classes under one accuracy claim; different conditions may need different inspection methods.
Treat lighting and presentation as part of the system
Assess camera placement, optics, lighting, fixturing, part motion and available line time. Preserve product identity and capture conditions in the evaluation set. If the feature is hidden, blurred or not distinguishable from acceptable variation, changing the model alone may not solve the inspection problem.
- Create a reviewed defect taxonomy with accepted parts, known defects and borderline cases.
- Include relevant batches, suppliers, finishes and changeovers; avoid a test set made only from ideal sample images.
- Record the line-rate and physical constraints, including how a part is identified and held for review after an uncertain result.
Connect a result to a controlled disposition
An inspection result needs the correct item or batch identity and a defined quality response. Show the evidence and uncertainty to the reviewer where required. Any reject gate, PLC handshake or MES record must be separately engineered and tested, including delays and failures. A classifier that works on stored images is not yet an accepted in-line inspection system.
Separate defect classification from metrology
Detecting a missing feature is different from measuring a dimension to a specified tolerance. Dimensional use requires appropriate optics, calibration, reference measurements and a measurement-system assessment. Do not infer metrology capability from image resolution or a model score. Existing gauges or rule-based vision may be the better choice for a stable, well-defined measurement.
Evaluate the station under representative conditions
Agree a reference disposition with quality engineers and hold back representative parts for evaluation. Report false accepts, false rejects and uncertain cases by defect class, product and operating condition. Test motion blur, dirty lenses, lighting variation and wrong product selection. Assess end-to-end timing, item association and failure behavior before considering a live reject function.
Define the quality and operating evidence
A scoped output may include an imaging review, labelled dataset, prototype, class-level test results and an integration specification. Handover identifies model and camera versions, reference images, monitoring, fallback and update ownership. New parts and newly observed defects require controlled review; the system should not retrain itself into production without an approved evaluation and release decision.
An illustrative assembly-presence check
A station is assessed for a missing fastener on one assembly family. Quality reviewers label representative views, including occlusion and acceptable variants, then compare model results with independent dispositions. Uncertain parts follow the existing review process. This is an illustrative evaluation, not evidence of zero defects, improved yield or validated dimensional measurement.
Bring parts, constraints and the quality owner
Start with the inspection specification, sample images or parts and the current process. Agree which errors matter and whether the necessary feature is visible before selecting an AI approach. Compare a candidate model with the existing inspection and simpler alternatives. Wider line coverage follows only when each station’s evidence and operating responsibilities justify it.
More clarity
Questions and answers
Will AI find every defect?
No. Performance depends on visibility, representative evidence and operating conditions. Unknown defects and changed presentation can cause misses, so limitations and fallback must be explicit.
Can the system run at our production line rate?
That must be tested with actual capture, processing, item tracking and response timing. A fast model benchmark alone does not establish end-to-end line suitability.
How many images are needed?
The evidence must cover the relevant product variation and defect classes. There is no universal sample count; rare defects and insufficiently represented conditions remain explicit gaps.
Can it replace a dimensional gauge?
Only a separately scoped and validated measurement workflow could support that decision. Visual classification does not establish measurement uncertainty or compliance with a tolerance specification.
How are product changeovers handled?
Product identity, camera configuration and model selection must be controlled and tested. A new product or setup requires review before previous acceptance evidence can be applied.
Start with one business problem