AI visual inspection puts industrial cameras and an edge computer on your line, classifies each part in milliseconds, and sends reject signals to your PLC — tuned to the false-reject rate your quality manager signs off, not a vendor default.
AI visual inspection is automated quality control that uses cameras and deep-learning models to find defects in products as they move down a production line. It outputs a pass, fail or review decision per part, the defect type and location on the image, and a signal to divert rejects. Unlike rule-based machine vision, it learns variable defects such as scratches, porosity or wrinkles from labelled examples.
Inspectors get tired, lighting varies, and sampling plans only look at a fraction of output. The result is escapes that reach customers and good parts thrown away out of caution.
of defective parts caught by trained visual inspectors of precision parts in a 2015 Human Factors study — the rest passed.[1]
of acceptable parts wrongly rejected by the same inspectors, a hidden scrap and rework cost.[1]
of gross sales lost to poor quality is the expert range Quality Digest cites for manufacturing and service firms.[3]
higher defect detection than human inspection is possible with AI image recognition, per McKinsey.[2]
Most inspection projects fail on optics, not models. We spec the camera, lens, lighting and trigger for your part and line speed first.
Models run on an edge computer beside the line, so decisions do not depend on the plant network or the cloud.
Every decision is logged with its image so quality can audit it and process engineers can see defects trend by cavity, shift or supplier lot.
We collect golden and defective samples, agree defect classes and limits with QA, and test lighting and camera options on the bench.
Station mounted in shadow mode. Images are captured at line speed and your inspectors label a few thousand parts with our tooling.
The model decides on every part without acting. We compare its calls with inspectors and a re-inspected audit set to measure escapes and false rejects.
If the agreed numbers are met, the PLC reject gate is switched on. You get the confusion matrix, image evidence and a rollout plan for other stations.
| Criterion | Manual inspection | Rule-based machine vision | Stratgik AI build + run |
|---|---|---|---|
| Coverage | Sampling or tiring 100% checks | 100% of parts | 100% of parts, with images kept |
| Variable defects (scratches, porosity, wrinkles) | Judged by eye, varies by shift | Hard to write rules for | Learned from labelled examples |
| Fixed checks (presence, dimensions, barcode) | Slow | Excellent and cheap | Supported, but rule-based tools may be enough |
| False-reject control | Inspector caution | Threshold tweaking | Operating point agreed with QA and monitored |
| New product or defect | Retrain people | Reprogram rules | Add labelled images and retrain |
| Traceability | Paper or tick sheets | Pass/fail logs | Image, decision and confidence per serial or lot |
Edge hardware now runs vision models at line speed for a modest cost. The hard part is choosing the right operating point and keeping it there as products and suppliers change.
Two leaks: good parts scrapped and bad parts reaching customers. The cut in each is an assumption; the shadow run measures your real numbers before the gate goes live.
Illustrative estimate using your inputs and stated assumptions, not a quote or guarantee. The pilot measures the real figure against your baseline.
One station, one product family, 30 days
Per station cluster, per month
Multi-line or multi-plant
Cameras, lenses, lighting, edge computers and mounting billed at cost with no markup; installation travel quoted separately; taxes excluded. GBP and AED prices are indicative conversions from USD.
Accuracy is measured as two numbers on your parts: the share of real defects caught and the share of good parts wrongly rejected. Both depend heavily on lighting, camera resolution and how clearly defects are defined. We report them from a shadow run against your inspectors and a re-inspected audit set, and set the operating point with your quality manager before rejects are automated.
False rejects are controlled by choosing a confidence threshold against a reject budget your QA team approves, and by adding a hold-for-review lane for borderline parts. Parts the model is unsure about are imaged and queued for a person instead of scrapped. Reviewed images feed retraining, so the borderline lane shrinks over time.
Traditional rule-based machine vision is the better choice for fixed checks such as presence, dimensions, barcodes and label position. AI visual inspection earns its cost on variable defects — scratches, porosity, surface texture, wrinkles or contamination — where writing rules breaks down. Many lines use both, and we will tell you if rules alone will do.
Usually hundreds of examples per common defect type, plus a larger set of good parts, is enough to start. Rare defects can be handled with anomaly detection trained mostly on good parts, then refined as real defects appear. The pilot's first two weeks are designed to collect and label these images on your own line.
A Stratgik pilot for one station and product family is $22,000 over about 30 days, with cameras, lighting and edge hardware billed at cost. Ongoing run is $4,500 a month per station cluster for monitoring and retraining. The return usually comes from fewer scrapped good parts and fewer customer escapes.
Inference runs on an edge computer at the station, so each decision arrives within the line's cycle time and keeps working if the plant network drops. Only images selected for review, retraining or audit are sent to your cloud account or on-premise storage. That also keeps production images inside your own environment.
We'll run a bench optics test and tell you, in plain numbers, whether AI inspection can hold your defect and false-reject targets before you commit to a pilot.