How to Choose AI Visual Inspection Equipment: A Selection Guide for Industrial Presence/Absence Inspection Equipment
Selection guide for industrial vision presence/absence inspection equipment: item by item, it explains how to determine object size, smallest detectable target, FOV, working distance, lens focal length, pixel sampling, light source illumination, algorithm selection and PLC communication interlocking.
When selecting AI visual inspection equipment, the key is not comparing specification sheets, but rather "Minimum defect size → required pixel resolution → field of view and working distance → lens and light source → algorithm and communication" work backward item by item in the order: first measure how large the smallest defect that must be detected is, then calculate the required pixel equivalent from that, and only then discuss camera and lens models. The pixel count only determines how many points there are in the image; whether those points can reflect a real difference depends on how the lens, light source, workpiece, mounting distance, mechanical stability, and algorithm perform together on site.
Clarify One Thing First
Pixel Count Is Not Inspection Accuracy
The most misleading point in selection is Treating "pixel count" as "inspection capability". The fact that the equipment has 5 MP does not mean it can measure a 0.01 mm error—the pixel count only determines how many points are in the image, and whether those points reflect real differences depends on the combined effect of the lens, light source, workpiece, mounting distance, mechanical stability, algorithm and site environment.
So this guide does not offer the shortcut of "choosing equipment from a specification table"; instead it Questions to confirm item by item during selection Laid out plainly: every question corresponds to a condition that must be implemented on site; miss one, and the inspection result may fall short of expectations.
selection Factors
Confirm Each Item One by One; Missing One Item Can Fall Short of Expectations
The following dimensions need to be clarified one by one during actual selection. Their order also reflects the sequence of confirmation: First determine what to inspect and how small it is, then work back to the optics and mounting conditions.
Inspection Object Dimensions
The overall size of the workpiece sets the starting point for the field of view (FOV)
Smallest Inspection Target
The smallest feature to be resolved determines the imaging resolution required
FOV (Field of View)
The area the lens can cover at a given working distance
Working Distance WD
The lens-to-workpiece distance is constrained by installation space and safety clearance
Lens Focal Length
The longer the focal length, the smaller the field of view and the stronger the sampling capability per unit area
Pixel Sampling
The real-world size each pixel represents is the upper limit of accuracy, not the accuracy itself
Light Source and Illumination
Making the target feature visible often decides success more than the algorithm does
Motion Speed and Cycle Time
Determines the usable exposure time and therefore the required brightness
FPS and Processing Capability
Acquisition and processing speed must match the production line cycle time
Algorithm Selection
Use rule-based algorithms where acceptance criteria are clear, and anomaly detection where they are hard to enumerate
PLC / I/O / Communication
How the results are fed into the existing production line control system
Installation Space
The space available for cameras, light sources and cables is often a hard project constraint
Field of View and Pixel Sampling
First Work Out How Many Pixels the Target Occupies in the Image
| Lens Focal Length | Working Distance | Field of View (H × V) | Pixel Sampling Reference Value |
|---|---|---|---|
| 8mm | WD 200mm | H 129.6 × V 97.2 mm | Approx. 0.050 mm/pixel |
| 16mm | WD 200mm | H 64.8 × V 48.6 mm | Approx. 0.025 mm/pixel |
| 25mm | WD 200mm | H 41.47 × V 31.10 mm | Approx. 0.016 mm/pixel |
The pattern is clear: Increasing the focal length narrows the field of view and raises the pixel sampling capability per unit area. This is why "changing a lens" directly changes the range the equipment can cover and the order of magnitude of detail that can be resolved.
light source and illumination
Make the Target Features Stand Out
In most practical projects, The illumination plan decides whether inspection can be done at all, more than the algorithm does. The same defect may be clearly visible or disappear completely under different illumination angles.
Reflective Parts
- Specular reflection can swamp features
- Common techniques: diffuse light, polarization, adjusting the angle of incidence
- The key is to keep reflected light out of the lens
Black Parts
- Light absorption causes low contrast
- Low-angle grazing light casts shadows at edges
- Where necessary, use specific wavelengths to enhance material differences
Transparent Parts
- Visibility depends on reflection and refraction
- Backlight or dark-field illumination is commonly used
- Batches with different transparency may behave inconsistently
Metal Parts
- Machining texture and reflections are both present
- Coaxial light suits flat surfaces
- Be careful not to create "false defects" with the lighting
Oil Stain and Contamination
- Oil film changes local reflection
- Use diffuse illumination to weaken the oil film
- If severe, cleaning before inspection should be considered
Ambient Light
- Workshop lighting varies
- Adding a light shield or increasing the light source share is recommended
- Scenes with direct natural light are the least stable
algorithm Selection
Choose by Acceptance Criteria Form, Not by How Advanced It Is
The principle for algorithm selection is simple: Use rule-based algorithms where acceptance criteria are clear, and AI only where they cannot be enumerated. Doing it the other way around only makes an otherwise stable project unexplainable.
template matching
targets that are fixed in position and stable in form. Simple to implement and interpretable, it is the most common type of "presence/absence inspection".
Anomaly Detection
occasions where defect types are difficult to enumerate. Modeling the normal state judges anything clearly deviating from the reference as an anomaly.
Classification
Where models, colors or states must be distinguished, output a category rather than presence/absence.
OCR Character Recognition
Reads printed text, inkjet codes, and engraved characters and verifies the content, suitable for traceability and error-proofing.
code reading
Barcode / QR code / DataMatrix decoding, with high robustness and fast speed.
Geometric Measurement
Where a dimension or position value must be output, such as seating depth, spacing or angle.
Communication and Interlocking
How results connect to the production line
The result ultimately must The control system that connects to the production line, so the communication method must be confirmed at the solution stage rather than after the equipment is built.
| Scenarios | Recommended Method | Description |
|---|---|---|
| Only one release / rejection signal required | I/O | Simple wiring, fast response, suitable for single-piece rejection |
| Needs to exchange status and specifications with the PLC | S7 / Modbus / Profinet | Can transmit model, recipe number and inspection results |
| Data must be returned to the host system | TCP | Suitable for integration with MES or traceability systems |
| Needs to interface with an older control system | RS485 | Low wiring cost, suitable for long distances and multiple nodes |
Another thing to confirm is Mechanical Stability the stiffness of the camera and light source mounting, the repeat positioning accuracy of the tooling, and light source decay over long-term operation. These factors do not appear in the specification sheet, yet they directly determine whether inspection is stable.
Common Question
Common questions about selection
Is 5 MP enough for presence/absence inspection?
It depends on How many pixels the target occupies in the image, rather than looking at the total pixel count. With the same 5 MP sensor, covering a 100 mm field of view gives about 0.02 mm per pixel, while covering a 500 mm field of view gives about 0.11 mm — the former can resolve fine features, the latter may not even show the target's contour clearly. The criterion should be "how many pixels the target occupies".
How do you choose a lens for visual inspection equipment?
The order is: first determine the field of view and working distance that must be covered, then work back to the focal length; next check whether the pixel sampling can resolve the smallest target to be inspected. A longer focal length is not automatically better — a smaller field of view means more cameras or a greater mounting distance.
Is pixel sampling value the inspection accuracy?
No. Pixel sampling describes "how much real size each pixel corresponds to", which is the theoretical upper limit of accuracy. Actual resolvable capability also depends on whether edges are sharp, whether contrast is sufficient, whether the light source is stable, and whether the mechanics are rigid. Treating the sampling value directly as an accuracy commitment is the most common mistake in selection.
Is an AI algorithm better than a traditional algorithm?
It is not "better," but "suited to different problems." For projects with clear criteria (presence/absence, count, characters), rule-based algorithms are more stable, faster and more explainable; only where the criteria are hard to enumerate (appearance anomalies) is AI anomaly detection the right fit. The two can also be combined.
What is most often overlooked during selection?
The most common answer: Installation space and mechanical stability. Many solutions work well in the laboratory but become unstable on the production line because of insufficient bracket stiffness, poor tooling repeat positioning, or light source aging. These factors should be included in the evaluation during the solution phase.
Submit sample testing
The most effective step in selection is to test actual samples under real optical conditions. Paper-based parameter calculations cannot replace measured results.
Submitted successfully
We have received your sample testing request. A solution engineer will contact you within 1 business day via contact you.
Need to assess imaging conditions, defect criteria and cycle time item by item? Go to the Full Requirement Assessment →
The Most Effective Step in Selection: Test Your Own Sample
For the same workpiece, the inspection result differs greatly with different lenses, light sources, mounting distances and algorithm combinations. Sending us samples for measurement is more reliable than extrapolating from a specification table.