Machine vision for quality control: where it can work today

When people talk about computer vision in manufacturing, the conversation quickly drifts to one of two extremes. Either it is "a revolution that will replace the entire quality inspection function", or it is "an expensive toy that does not work under real conditions." Neither is accurate.

I look at machine vision as a tool with specific limits of applicability. Where a task falls within those limits - it works well and produces measurable results. Where it does not - no amount of investment will fix that.

Which tasks are a good fit

Machine vision handles tasks well when they:

• have a clear and repeatable visual criterion - size, shape, part presence, colour, a crack or scratch;
• run under controlled conditions - consistent lighting, a fixed camera angle, stable conveyor speed;
• require a checking frequency that a human physically cannot sustain without errors;
• have clear economics - the cost of a missed defect exceeds the cost of the system over a reasonable period.

Typical examples: verifying assembly completeness (are all components present), checking geometric tolerances, detecting surface defects on uniform parts, inspecting markings and labels.

Where it does not work

It is not worth expecting good results when:

• the quality criterion is subjective or shifts from batch to batch;
• lighting conditions are unstable or the product surface reflects differently each time;
• defects show up not visually but in material properties or product behavior;
• the sample of defective parts available for training the system is too small.

A good machine vision system is not "we installed a camera and it figures things out." It is a clear specification, controlled capture conditions, and a set of reference samples - both good parts and defects.

How to evaluate the economics before implementation

Before looking at vendor proposals, answer a few straightforward questions:

1. What is the current defect rate at this production stage?
2. What does one missed defect cost - including returns, rework, and reputation?
3. What does manual inspection at this stage cost per year?
4. What share of defects are actually visual and detectable on the line?

If the combined losses from defects and manual inspection significantly exceed the cost of the system and its maintenance over three to four years - the economics work. If not, you either need to find a different production stage, or honestly acknowledge that this task is not a good fit.

What the production side needs to provide

Deploying a machine vision system is not purely about technology. You also need to:

• identify a production area with stable conditions, or invest in creating them;
• assemble a training dataset: several hundred good parts and an equal number of defective ones covering the relevant defect types;
• define who maintains the system and what happens when it produces false positives;
• integrate the system's signal into the production process - stopping the line, rejecting the part, alerting the operator.

The technology itself is accessible today and is not the barrier. The barrier is usually an unprepared production environment and the absence of a clear process for acting on the system's output.

A practical filter

Before talking to a vendor, ask yourself three questions:

1. Can I show on concrete data what this defect category costs per year?
2. Can I guarantee stable capture conditions at this production stage?
3. Do I have a set of reference samples - both good and defective?

If all three answers are yes - a conversation about the system makes sense. If not - solve those first. Otherwise the system will be installed but will not perform as expected.