Machine Vision Inspection Setup: A Quality Manager's 8-Step Checklist

If you're setting up machine vision inspection on a production line, this checklist is for you. I'm a quality compliance manager at a mid-size contract electronics manufacturer, and I review every inspection system before it reaches our floor—roughly 35–40 deployments per year. I've rejected 18% of first-time setups in 2024 because they missed verification basics. This is the eight-step checklist I actually use.

It's built around Cognex systems because that's what we've standardized on, but the process applies to any vision platform. Follow it in order and you'll avoid the install mistakes I've watched teams repeat—including a few I've made myself.

Step 1: Define what "defect" means on your line

You can't configure a vision system to catch a defect you haven't written down. Sounds obvious, but I reviewed two deployments last year where the requirement was verbal. "Catch anything that looks wrong." That's not a spec; it's a hope.

Sit with your process engineers and document:

• Each defect type you're checking for (missing components, contamination, mislabeling, cosmetic scratches, dimensional drift)
• Acceptable tolerance for each defect type
• Where on the part or packaging the defect appears

Then do a reality check: can a camera actually see it? If the defect is internal—like a void in a molded part—a visible-light camera won't catch it. You'll need a different technology (more in Step 2).

Step 2: Match the technology to the defect type

This is where I see the most money wasted. Teams buy a machine vision system because it's the standard, then discover it can't detect their actual failure mode.

Here's how I split inspection tasks:

• Presence/absence, surface defects, OCR, barcode grading → Cognex machine vision inspection (In-Sight cameras or VisionPro)
• Heat-related failures (weld penetration, overheated components, insulation gaps) → Thermal cameras, like FLIR models
• Tight dimensional tolerances and GD&T callouts → Coordinate measuring machine, like a Zeiss Contura

Real example: in late 2024, we quoted a weld inspection station for an automotive supplier. The customer insisted on a standard vision camera. But the rejection mode was insufficient weld penetration—that's a heat signature issue, not a visible surface issue. We ended up pairing a Cognex In-Sight camera for weld position verification and a FLIR thermal camera for penetration quality. The thermal camera caught defects the vision camera literally couldn't see.

If you're new to FLIR cameras, here's the short version of how to use one on a production line:

1. Set emissivity first. Most machined metals read around 0.15–0.3 without correction, which produces wildly wrong temperatures. Apply high-temp tape or paint to the target to push emissivity close to 0.95, or enter your material's value directly into the camera.
2. Focus manually. Autofocus drifts on small, high-speed parts.
3. Mount it fixed. Handheld thermal cameras don't give repeatable readings on a moving line. Consistent distance and angle matter more than raw resolution.
4. Calibrate at every shift start. Thermal drift is real. We check against a known reference surface before running.

And for the Contura CMM: use it for first-article inspection and periodic sampling, not inline 100% checking. CMMs are precise but slow. They complement a vision system; they don't replace it.

Step 3: Use the Cognex downloads before you commit to hardware

This is a free step that most teams skip. Actually, let me be more precise: they skip it until after they've bought, then find out the software workflow doesn't fit their process.

Go to the Cognex downloads area and get the software for the product family you're considering. For In-Sight cameras, that's In-Sight Explorer. For PC-based vision, it's VisionPro. There are also sample programs for common applications—presence/absence, OCR, barcode reading.

Run a sample. Play with the thresholds. See what the interface expects.

Cognex updates its software regularly, so verify what's current as of your project date. I'm writing this in early 2025, and the workflow has been consistent across the last several releases—but don't skip the version check.

Step 4: Set up lighting before you touch the camera lens

Lighting is the hidden killer of machine vision projects. I've seen systems that pass in the lab, then go blind on the factory floor because skylights change the ambient light at 3 PM.

Rules I follow now:

• Use controlled light: backlight for silhouettes, ring light for surface inspection, bar lights for barcode reading.
• Enclose the inspection station as much as possible to block ambient light.
• Don't compensate with exposure time. Long exposures blur moving parts.

I once skipped the lighting enclosure because "what are the odds the environment changes?" Well, the odds caught up with me when the system false-rejected 30% of parts on a cloudy afternoon. The trigger sensor was picking up IR noise from a nearby furnace door. So glad we rebuilt that station with a proper enclosure and dedicated lighting. Almost went with cheap LED strips to save $300—they flickered at 60 Hz and wrecked the image quality. The $1,100 proper lighting kit was the right purchase.

Step 5: Test with known-good AND known-bad parts

Here's the step I see skipped about 70% of the time: teams test with known-good parts, tune the thresholds to pass, and assume it's done. They never feed actual defective parts through the system.

My minimum protocol:

1. Collect 20–30 known-good parts that include normal production variation.
2. Collect 10–20 known-bad parts (from past failures, or deliberately created).
3. Run both sets. Target: 100% pass on good, 100% reject on bad.
4. If you can't hit both, adjust lighting or thresholds before you move on.

This caught a serious issue for us in Q1 2024. The known-bad set revealed that a subtle oil residue defect wasn't being detected at line speed. If we'd gone live without that test, we'd have shipped contaminated PCBs and eaten a six-figure recall cost.

Step 6: Decide what happens AFTER a rejection

People forget this constantly. When the vision system rejects a part, what's the physical action?

• Air-jet diverter? Stop barrier? Conveyor stop?
• Operator notification via HMI, alarm, or text?
• Image saved for review? Traceability record?

If you don't plan the rejection workflow, you get chaos. I saw a plant in late 2024 where the vision system was doing its job—rejecting a misaligned label—but the operator silenced the alarm because nobody had trained him on the reject protocol. Rejected parts piled up in a bin, and then the bin got emptied back into the good stream by mistake. 8,000 units quarantined, all because the workflow wasn't designed.

The most frustrating part: the vision system was flawless. You'd think an alarm means "stop and check," but if an operator has been burned by false alarms before, they'll tune it out. That's on the team that installed the system, not the operator.

I now require a written rejection-response protocol on every install: one line for what the system does, two for who responds, three for what they do with the part, four for when to escalate.

Step 7: Test the full line integration (even when you're on a deadline)

This is where time pressure gets dangerous. The line is scheduled to start Monday, the camera arrived Thursday, and you're tempted to test the camera in isolation and trust the integration.

Don't.

Run the full cycle: part enters → camera triggers → image captures → judgment returns → reject action executes → data logs. Do it 50 times. Then pull the logs and inspect them.

In March 2024, we paid a contractor about $400 extra for a full day of integration testing on a rush project. The customer's launch was tied to a $15,000 event, and missing it wasn't an option. During that test we found a PLC handshake issue: the position trigger fired 120 ms too early, so every image was captured mid-motion and blurred. We fixed it before launch and passed the customer's installation review on the first attempt.

When the deadline matters, paying for a verified, predictable install beats hoping for the best with a cheaper one.

After getting burned twice by "probably on time" schedules in earlier years, I now budget for integration testing on every project. It's non-negotiable.

Step 8: Document everything and train the operators

The system isn't finished when it passes 50 test cycles. It's finished when the operators can run it and the next engineer can troubleshoot it.

Minimum documentation pack:

• Which defect types the system detects, and why those specific ones
• Threshold settings and the rationale behind them
• How to access the image and rejection logs
• Escalation contact for your integrator or Cognex support

Then train operators on the rejection-response tree. Not video training—hands-on. Walk through a real reject, a false alarm, and what the HMI is telling them.

Common mistakes I still see after 4 years of reviewing vision installs

Skipping the known-bad test. I said it in Step 5, and it's worth repeating. If you haven't proven the system rejects actual defects, you haven't proven anything.

Choosing the cheapest vendor when you're on a deadline. I've seen three plants in the last 18 months pick a lower-cost vision vendor to save $5,000, then burn two months fighting installation issues. The delays cost more than the savings. When deadlines matter, the reliability of a proven system—and the support behind it—is worth the premium.

Assuming one tool inspects everything. Keep thermal cameras for heat-related failures and a CMM for precision dimensional sampling. Each tool has its lane.

Trusting default settings. Cognex sample programs are a starting point. Your lighting, part variation, and line speed are different from anyone else's. If you don't tune and test, you're just pointing a camera at a conveyor and hoping.

This checklist reflects what I knew as of Q1 2025. Vision technology moves quickly—Cognex and other vendors release new models and software versions all the time—so verify current specs, pricing, and support options before you budget.