# Visual quality inspection

> Visual quality inspection means imaging products on the line with cameras and using an AI model to detect defects such as scratches, cracks, missing parts or colour differences. This page describes Corrplus's approach rather than a client case study: we first measure whether your data and line suit it, then trial it at a single station.

URL: https://corrplus.net/solutions/visual-quality-inspection/
Updated: 2026-10-22
Corrplus Team

## Symptoms

- Inspection is done by eye, and results vary by shift, person and fatigue.
- As the line speeds up, not every item is checked and sampling has to do.
- Defects are noticed only when a customer returns the product.
- There is no record of which defect type comes from which machine or shift.
- When an experienced inspector leaves, their knowledge leaves too.

## What waiting costs

Every defect that reaches a customer means a return, a dent in your reputation or a contract at risk. Without a defect record, finding the root cause stays guesswork.

## Our approach

- **Feasibility**: We collect images of good and defective products from your line and test whether the defect types can be told apart on camera.
- **Imaging setup and labelling**: We fix the camera, lighting and trigger setup, and build defect definitions and a labelled image set with your quality team.
- **Single-station pilot**: The model first runs in alert-only mode, and its calls are compared with your inspectors' decisions.
- **Connect to the line**: If results hold up, the model is connected to the reject mechanism, PLC or MES, and every detection is logged with its image.
- **Monitor and retrain**: New products, new defect types or lighting changes trigger retraining with fresh images.

## What to measure

- Defect escape rate to customers
- False reject rate on good products
- Share of items inspected
- Traceability by defect type, machine and shift

## Does computer vision work on every line?

No, and it is worth saying up front. A defect can only be detected by camera if it shows in the image. Surface scratches, missing screws or wrong labels usually do. A void inside the material or a tiny dimensional deviation may need other sensors.

The second condition is consistency. Camera angle, lighting and product position must stay the same in every image. Daylight from a window or a vibrating conveyor can undo everything the model has learned. So the first half of a project is often more about the physical layout of the station than about software.

The third is labelled data. A model can only learn what your quality team calls a defect from examples. If inspectors on the same team judge the same part differently, defect definitions have to be agreed first.

## Which setup option fits you?

| Option | When it fits | Data needed | Flexibility |
|---|---|---|---|
| Off-the-shelf industrial vision system | Standard measurement, presence checks, barcodes | Low | Limited to the vendor's tools |
| Cloud vision service (Azure, Google) | Quick trials, slower lines | Medium | Medium, needs connectivity |
| Custom model on an edge device | Complex defects, fast lines, data must stay on site | High | High, needs upkeep |

The right option depends on defect type, line speed and whether data can leave the plant. Sometimes the right answer is an off-the-shelf system, and we will tell you so.

## How we start

The first step is feasibility within an [AI consulting](/services/ai-consulting/) engagement: we collect images from your line, define defect types with your quality team and measure whether detection is possible. If it is, we run a pilot at one station. The model starts by raising alerts only; it does not stop the line.

During the pilot we log every call the model makes next to your inspector's decision. We watch two measures: defects the model misses and good parts it rejects. We agree acceptable limits for both with you, based on the risk of your product. If the limits are not met, we do not connect the model to the line; we report why and what the options are.

Training runs on [cloud infrastructure](/services/cloud-infrastructure-scalable-platforms/), and inference on the line runs on an edge device. After go-live, monitoring and retraining with new images continue under [managed services](/services/managed-services/).

## Frequently asked questions

### How many images do we need?

It depends on the defect type and imaging conditions, so we do not quote a fixed number. You need enough labelled examples of each defect and many images of good parts. Rare defects are the hardest part; feasibility measures this.

### Are our existing cameras good enough?

Security cameras usually are not. Depending on defect size, line speed and surface, you need an industrial camera with suitable resolution and controlled lighting. During feasibility we test your current equipment and list any gaps in writing.

### Why a custom model when off-the-shelf systems exist?

Off-the-shelf industrial vision systems are strong for simple, standard checks. When defect definitions are complex, product variety is high or integration needs are specific, a custom model is more flexible. We compare the options against your needs.

### Will it replace our inspectors?

No. The model takes over the repetitive looking; doubtful cases, new defect types and final decisions stay with your inspectors. Their knowledge goes into the model through labelling.

### What if the cameras capture workers?

Cameras are positioned to see the product, not people, wherever possible. If workers appear in images, that is personal data under GDPR and Türkiye's data protection law, KVKK; notices, retention and access rules are agreed with your legal counsel.