Most manufacturers I talk to are curious about computer vision for quality control, but they’re not sure where to start. The good news? You don’t need a PhD or a million-dollar budget. Here’s what actually works on Australian factory floors.

Start with Your Camera Setup

First things first: you need decent cameras. Don’t overthink this initially. A $500-800 industrial USB3 camera with a resolution around 5MP will handle most inspection tasks. Brands like Basler or FLIR make reliable units that can survive the dust and vibration of a production line.

Mount your cameras perpendicular to the inspection surface, about 30-50cm away depending on the part size. Fixed mounts are crucial – even a millimetre of camera shake will ruin your accuracy. I’ve seen factories waste weeks troubleshooting “model problems” that were actually just loose camera brackets.

Position matters more than most people realize. You’re typically looking at a field of view between 100-300mm wide, so plan your camera placement before you drill anything. Mock it up with a cheap webcam first.

Lighting Makes or Breaks Everything

Here’s what nobody tells you: lighting is 60% of your success. Computer vision models can’t fix shadows, glare, or inconsistent illumination.

Use LED ring lights or diffused panel lights positioned at 45-degree angles to minimize shadows. For around $200-400 per light, you can get industrial-grade LED panels with consistent colour temperature (usually 5000K daylight).

The key is consistency. Your lighting needs to be identical at 8am and 6pm, whether it’s sunny or cloudy outside. Enclose your inspection area if possible, or at minimum block direct sunlight from hitting your parts.

For metallic surfaces, diffused lighting is essential. For dark materials, you might need backlighting. Test different setups with your actual parts before committing to a configuration.

Training Your Model with Real Defects

This is where it gets interesting. You need defect samples to train your model – and you need more than you think.

For a basic pass/fail inspection, aim for at least 200-300 images of good parts and 200-300 images of defective parts. More is better. Variation matters too: different angles, slight position changes, various defect severities.

Don’t have enough defects? You’ll need to create them. Yes, really. Controlled defects that match real-world failures. It feels wrong to intentionally scratch parts or create voids, but you can’t train a model on problems it’s never seen.

Label your images carefully. Most platforms use bounding boxes or pixel-level segmentation. This is tedious work – budget 30-60 seconds per image. For 500 images, that’s 4-8 hours of someone’s time. There’s no shortcut here.

Pre-trained models exist (check out platforms like Roboflow or Landing AI), but you’ll still need to fine-tune them with your specific parts and defects. Generic “scratch detection” won’t catch the particular flaws in your manufacturing process.

Integrating with Your PLC Systems

Now for the practical bit: connecting your vision system to your existing production line. Most factories in Australia run on PLCs from Siemens, Allen-Bradley, or Schneider Electric.

You’ve got two main options:

Option 1: Digital I/O signals. Your vision computer sends a simple pass/fail signal to your PLC via digital outputs. This is the simplest approach. A relay module costs $100-200 and your electrician can wire it in a few hours.

Option 2: Industrial protocols. If you need to send detailed defect data (type, location, severity), you’ll want Modbus TCP or EtherNet/IP communication. This requires more setup but gives you much better data for tracking trends.

Start with Option 1. Get the basic system working, then add complexity if you need it.

Your cycle time matters. If your parts move past the camera in under 2 seconds, your vision system needs to capture, process, and send a signal in that window. Modern edge computers can handle inference in 50-200ms, but factor in your camera exposure time and any PLC delays.

What Accuracy Should You Actually Expect?

Let’s be realistic. A well-configured computer vision system typically hits 95-98% accuracy on clear, well-defined defects like cracks, chips, or dimensional issues.

That’s not perfect. You’ll get false positives (rejecting good parts) and false negatives (missing bad parts). Plan for this. Most factories run vision inspection alongside manual spot-checks, at least initially.

For critical defects that could cause safety issues, don’t rely on vision alone. Use it as a first-pass filter, not your only defence.

Subtle defects are harder. Surface discolouration, minor texture changes, or hairline cracks might only get you to 85-90% accuracy. You’ll need more training data and possibly better cameras.

The CSIRO has published some excellent case studies showing real-world accuracy rates from Australian manufacturers. Their data suggests that most vision systems need 2-3 months of tuning to hit their target performance.

Start Small, Then Scale

My advice? Pick one inspection task. One part, one defect type. Get that working reliably before you expand.

A single-camera system with decent lighting and a focused model can be running in 4-6 weeks for under $10,000 in hardware. That’s cheap enough to prove value before you commit to a facility-wide rollout.

Computer vision for quality control isn’t magic, but it’s increasingly practical for everyday manufacturing. The technology’s there. The question is whether you’re ready to put in the setup time to make it work.