What can machine vision detect, and what can it not see?

What can a camera see?

A camera records the light coming from the surfaces facing it. Anything that changes that light can be detected: whether an item is there, where it sits, what is printed on it, what a code says, its outline and size, its colour and marks on its surface.

It sees one side per view. A round bottle with a label all the way round, or a part with several faces that matter, needs more cameras, mirrors or a way of turning the product. Clear containers are the useful exception: with a light behind them, the camera sees through the wall to the liquid line, which is how fill level is checked on PET and glass.

Why does contrast decide what is detectable?

A defect is detectable only if it sends the camera a different amount or colour of light than the good surface around it. That difference is contrast. A black smudge on a white label is easy. A shallow scratch on a shiny can may be invisible under one light and bright under another.

Shiny surfaces behave like mirrors: light leaves at the same angle as it arrived (HyperPhysics, law of reflection). So the angle of the light, not the camera, often decides whether a surface looks bright or dark, and whether a scratch stands out. Size matters as well: the defect must cover several pixels at the chosen field of view, or it disappears into the image.

How does lighting change what is visible?

Lighting is chosen for the defect, on your samples:

• Backlight turns outlines, holes and fill levels into sharp silhouettes.
• Low-angle (dark-field) light makes scratches, dents and embossed text glow on a dark surface.
• A dome evens out glare on shiny, curved packs so print and codes can be read.
• Coaxial light suits flat, glossy faces.

The colour of the light matters too. HIKROBOT’s lighting catalogue notes that the fluorescence of ultraviolet light can be used to detect UV glue and invisible codes (HIKROBOT Light Source Catalog 2025 Q3, p. 80). A shroud keeps daylight out so the station’s light is the only one that counts. The machine vision lighting guide covers each technique and its limits.

What can a camera not see, and what is used instead?

Some quality questions have no visible answer. The table sets out the common ones; the last column is the method that answers them, often on the same line as a camera.

addaScan builds camera stations. Where a question needs X-ray, leak testing or weighing, we say so early; that equipment comes from its specialist suppliers. Camera, X-ray or high-frequency fill inspection compares the fill methods.

How do you inspect something the pack will hide?

Inspect it before it is hidden. The foil seal is checked before the cap goes on, the insert before the lid closes, the items in a tray before it is wrapped, and the empty container before it is filled. That is one reason a line may have more than one station; chapter 4 covers placement.

Whether a fault is visible on your product is not decided from a datasheet. A sample test images your good and defective samples, including borderline ones, under candidate cameras, lenses and lights, and records what can and cannot be seen. The sample preparation guide lists what to send.

Test visibility before choosing an algorithm

The smallest defect is not always the hardest defect. A large transparent fold can create less contrast than a tiny dark particle. Contrast comes from how light interacts with the material: backlighting reveals a silhouette, low-angle light reveals changes in surface slope, and diffuse light can reduce glare. If the defect and acceptable product produce indistinguishable images in the selected view, a more complex algorithm cannot recover information that was never captured.

Worked example — hypothetical: A hypothetical pouch has a visible wrinkle across its seal region. The vision system can identify that appearance when the region is exposed and the image is suitable. A second pouch can look smooth while carrying a leak path below the optical detection limit. The correct requirement is therefore visible-seal-defect inspection, with integrity tested by the separately chosen method.

ASTM F88/F88M-23 scope 1.1 concerns seal strength in flexible barrier materials; ASTM F2096 concerns gross leaks detected by internal pressurisation. Neither public scope makes a visual pass proof of package integrity. (HIKROBOT Light Source Catalog, 2025 Q3, glossary) (ASTM F88/F88M-23: seal strength of flexible barrier materials) (ASTM F2096-11(2019): gross leaks by internal pressurisation (bubble test))

Keep the original image together with the active recipe and the reference judgement when investigating a disagreement. Change one controlled setting at a time and re-run the relevant challenge samples. Once a correction works, verify the neighbouring requirements too: improving contrast for one feature can hide another. Record the revised configuration and the evidence for release so maintenance does not have to reconstruct the decision from memory.