Computer VisionCase: Real-Time Surface Defect Detection on Metal Sheets Moving at up to 5 m/s

Table of contents

• The Production Flow
• What the Numbers Show
• Why It Matters for Quality and Production Teams
• Where This Leads

The challenge was not simply detecting defects, it was detecting them before defective material reached the next processing stage.

In cold rolling production, metal sheets move fast, and surface defects — holes, tears, and other irregularities — can appear at any point during high-speed processing. By the time a defect is caught manually, if it’s caught at all, the sheet may have already moved on to the next stage, carrying the problem with it. The result is a familiar chain of consequences: quality issues downstream, wasted material, unplanned downtime, and in the worst cases, defective product reaching a customer.

This case looks at how a computer vision system was built to catch these defects in real time, on a line moving at up to 5 meters per second, and route that information to the people and systems that need to act on it before the material moves further down the process.

The Production Flow

Rather than treating defect detection as an isolated inspection step, the system was designed around the actual sequence of events on the line: the sheet is detected, a defect is identified, its coordinates and characteristics are captured, the operator is notified, and the data is handed off to production systems. Each stage feeds the next, and the entire chain has to complete before the material reaches the next stage of processing — which, at 5 m/s, leaves very little room for delay.

1. Sheet Detection

The first task is establishing exactly where the sheet is in the camera’s field of view. The system continuously monitors the line and automatically detects the boundaries of the metal sheet as it moves through the frame. This isn’t a one-time calibration — the boundary detection runs continuously, frame by frame, so the system always knows precisely what portion of the image is sheet and what is background, machinery, or reflection. That distinction matters, because everything downstream depends on it: a defect can only be measured and located correctly if the system first has an accurate read on the material itself.

2. Defect Identification

With the sheet boundaries established, the system scans the moving surface for irregularities in real time. It’s built to recognize up to 10 distinct defect types, from holes and tears to other surface-level anomalies that indicate a quality problem. Because the line doesn’t slow down for inspection, this identification step has to keep pace with the material — there’s no second pass, and no opportunity to review footage after the fact. Detection accuracy for the system sits at up to 85%, a figure that reflects the difficulty of the environment: reflective metal surfaces, variable lighting, and fast motion all work against consistent detection, which is why visual conditions on the line are treated as a first-class part of the deployment, not an afterthought.

3. Coordinates, Size, and Classification

Detecting that something is wrong is only half the job. Once a defect is identified, the system fixes its exact coordinates on the sheet, estimates its size, and assigns it a classification based on the defect type. This is the layer that turns a raw detection into something actionable — a coordinate and a size are what let a downstream process decide whether to flag a section for rework, trim it, or reject it outright, rather than treating every anomaly the same way regardless of severity.

4. Operator Notification

From there, the information reaches the people who can act on it. The system sends timely notifications to the operator’s panel, alerting personnel before the defective section moves further along the line. This is the step that directly answers the original challenge: it’s not enough to log that a defect occurred somewhere in the run, the alert has to arrive while there’s still a chance to intervene — while the material is still upstream of the next processing stage, not after.

5. Production-System Handoff

Finally, the same data — coordinates, size, classification — is transferred to production IT systems, not just displayed on a screen. This closes the loop between detection and the systems that track quality, plan downstream processing, and maintain a record of what happened and where. It means the defect data doesn’t live only in the moment it was caught; it becomes part of the production record that quality and operations teams can use for trend analysis, supplier conversations, or process adjustments.

What the Numbers Show

Four figures anchor this case:

Up to 5 m/s — the processing speed the system operates at without requiring the line to slow down for inspection.

Up to 10 defect types — the range of surface irregularities the system is built to distinguish between, rather than flagging every anomaly as a single generic “defect” event.

Up to 85% detection accuracy — a realistic figure for a high-speed, reflective-surface environment, and one that improves as camera positioning and lighting are tuned to the specific line.

Real-time — the defining constraint of the whole system. Every stage, from sheet detection through to the operator notification and system handoff, happens fast enough to matter before the material moves on.

Seeing It in Context

On the real line, this looks like continuous overhead footage of the sheet moving through the camera’s field of view, with defect regions highlighted as they’re identified — holes, tears, or other flagged irregularities appearing as bounding boxes the moment the system catches them, tied to a coordinate and a defect class. Close-up stills of individual defect types, alongside a simple data-output view showing the fields the system actually generates for each event (coordinates, size, classification, timestamp), give a clearer picture of what operators and quality teams see in practice, beyond the high-speed overview shot.

Why It Matters for Quality and Production Teams

For a Quality Director, the value isn’t just automated detection — it’s a consistent, documented record of what was caught, where, and how severe it was, replacing a process that otherwise depends on catching problems by eye at speed. For a Production Director or Plant Manager, the real gain is avoiding the cascade: a defect caught at the point of occurrence doesn’t turn into downtime, rework, or scrapped downstream material. And for a CIO, the fact that defect data flows directly into production IT systems means it’s not a standalone tool generating alerts in isolation — it’s a data source that plugs into the systems already tracking production.

Where This Leads

If your line is running fast enough that manual inspection can’t reliably catch surface defects before they move downstream, the question worth answering isn’t whether computer vision can help in principle — it’s whether it can work on your specific line, your lighting, and your material.

We offer a defect-detection feasibility assessment using your own video samples: real footage from your line, evaluated against the defect types you actually see, to determine what detection accuracy and response time are realistic before any commitment to a full deployment.