Computer vision application
Automated visual inspection
Continuous inspection by camera and computer vision model, in place of manual sampling. This page covers what can be inspected, what the image needs, and where the limits are.
Tell us what needs inspecting. We'll assess whether the image can support the criterion.
What can be inspected
Categories of check that recur across sectors, even when the product changes.
-
Presence and absence
A component, seal, label, screw or item that should be there — and isn't.
-
Integrity and surface defects
Cracks, dents, scratches, stains, welding or finishing faults.
-
Positioning and assembly
A part in the wrong orientation, misaligned, or assembled out of sequence.
-
Labeling and print
A label that is wrong, crooked, unreadable, or carries the wrong date and batch.
-
Fill and conformity
Fill level, count per package and conformity with the batch standard.
-
Quality grading
Sorting by grade, size or category from visual criteria.
Image requirements
What has to be settled at capture before any discussion about the model.
-
Resolution where it counts
What matters isn't total megapixels, but how many pixels cover the smallest defect that must be detected.
-
Controlled lighting
Stable lighting solves more detection problems than a bigger model. Glare, shadow and varying light are common causes of error.
-
Position and angle
Defects that only show at one angle require that angle. Sometimes moving the camera is cheaper than training more.
-
Line cadence
Conveyor speed drives exposure time, the need for synchronized triggering, and the processing budget per part.
-
Environment
Humidity, dust, vibration, washdown and temperature determine camera protection before any software choice.
False positives and false negatives
Both errors always exist. What changes between projects is which one you tolerate more.
-
Each costs differently
A false positive discards good parts and erodes operator trust. A false negative lets the defect through. The balance is a business decision, not a technical one.
-
The threshold is adjustable
The model's cutoff can be shifted toward sensitivity or precision, depending on what each error costs on your line.
-
Human review where it fits
Borderline cases can be routed to human review instead of being classified automatically.
-
Measured against current inspection
The reference is the process you run today, not an isolated absolute number.
When automated inspection isn't the answer
Saying so early is cheaper for both sides.
-
The defect isn't visible
Internal, chemical or functional faults need a different kind of sensor.
-
No reference standard
If two conforming parts look very different from each other, there is no stable criterion to learn.
-
Very low volume
A handful of parts per day rarely justifies the cost of capture, installation and maintenance.
-
Subjective, shifting criteria
When two human inspectors often disagree, the first job is defining the criterion — not installing a camera.
Frequently asked questions about automated visual inspection
Technical questions that come up before assessing a line.
Does it work on a high-speed line?
It depends on the combination of cadence, exposure time and defect size. It's one of the first things checked in the feasibility analysis.
Does it replace the human inspector?
In practice it changes their role: the system covers every part repetitively, and the person handles exceptions and borderline cases.
Do I need special lighting?
Often yes, and it tends to be the highest-return investment in the project. Stable lighting reduces the difficulty of the problem before any model.
How is the inspection recorded?
Every event can be stored with image, date, time and classification, which makes the result auditable afterwards.
Talk to our team
Let's understand your challenge
Prefer to book the diagnosis conversation? Use the scheduling link. Or fill in the form — our technical team replies.