Inline panel quality inspection

Depro Profiles: bespoke profiles and panels

Depro Profiles is a Belgian family business based in Comines. Since 1992, the company has designed and manufactured high-quality profiles and panels for interior fit-out and the do-it-yourself market. Product diversification is central to its approach: from dimensions and finish to decor development. In a production process where every panel can be unique, consistent quality is one of the most important pillars.

The challenge

At 40,000 panels per day, that is a substantial volume to assess visually. Until recently, quality control was carried out manually by operators along the line. That approach had structural limitations:

• Subjective and tiring work: visually inspecting tens of thousands of panels a day inevitably means defects get missed, especially towards the end of a shift.
• No 100% inspection: at that production speed, it is physically impossible to examine every panel with the same level of thoroughness.
• No traceability: in the event of a complaint, there is no objective record of how the panel left the line.
• Differences between decors: Depro produces in many colors and wood types. What stands out on a light decor disappears on a dark one.

Note: the screenshots shown use illustrative demo data and do not reflect actual production data from Depro.

Approach: from lab test to dual inspection

A vision system on a production line isn't built in one go. Which camera reliably detects this type of strip on this type of panel? Which AI model recognizes the defect, even on a dark decor? That's why Depro and Aionix opted for a phased approach: start small, then scale up. Every phase produced lessons that we carried forward into the next.

Our solution

We built an inline AI vision system that inspects both sides of every panel for the quality of the glued decorative strips. The solution combines our own mechanical design, a custom-built electrical cabinet, an in-house trained AI model, and a cloud platform for traceability and insight.

On the platform this runs as the Quality module, with Traceability for the per-panel image archive.

Hardware: from 3D model to commissioning

The inspection had to happen inline, without interrupting the production line. Aionix designed a 3D model of an enclosure that groups camera, lens and lighting into one compact module, mounted along the line at the exact height and angle of the panel side. The design was manufactured externally, but Aionix handled the installation and alignment on the line itself. A deviation of just a few millimeters and the AI model receives images it wasn't trained for.

The electrical cabinet was also built by a specialized partner based on our specifications. Aionix then took care of the testing and commissioning on the line: all I/O signals, and the communication between the camera, the PLC and the vision platform.

AI model: trained in collaboration with the quality department

The AI model is the heart of the system. We trained it in close collaboration with Depro's quality department: they identified which images were good and which were bad, and we fed that knowledge into the model.

Three principles guided the approach:

• Training on bad samples. An AI model that has only ever seen “good” cannot recognize defects. We actively sought out examples of poor bonding and included them in the training set.
• Iterative validation with quality. The quality team reviewed the results after every round. Incorrect assessments were fed back into the next training round.
• Robust across all decors. The model was tested until it performed stably across the full product mix, and was fine-tuned on the combined dataset when the second camera was added in phase 3.

Image 1 of 4

100% inline inspection of every panel

Every panel that passes over the line is inspected on both sides. At 40,000 panels a day, that means 80,000 AI inspections per day: a volume that is visually impossible to achieve, and one the system carries out silently.

For every panel, the model produces a classification in milliseconds: approved or rejected. That result is stored immediately, together with the image itself. The physical removal of rejected panels currently still rests with the operators, but the foundation for further automation has been laid.

Cloud storage & full traceability

For every panel inspected, both the image and the classification are stored in the cloud. In the event of a customer complaint, Depro can look up the exact panel and see how it was classified at the time of production. Instead of a gut feeling of “I think it was fine,” there is now an objective record.

On top of the archive runs a dashboard with live trends: rejection rates by decor, shift and machine. Quality issues are tackled at the source, rather than piling up on the complaints desk.

Results

With the AI vision system in production, quality control at Depro looks fundamentally different from before:

• From sample-based visual inspection to 100% inline AI inspection of every panel, on both sides.
• 80,000 AI inspections per day, a scale that is not achievable manually.
• Full traceability of every panel via the cloud archive.
• Live insight into quality trends, and an AI model that performs stably across all decors.

What's next?

The current setup already delivers the objective data and traceability Depro was looking for. But the potential of a vision system on a line producing 40,000 panels a day goes further. Two next steps are on the roadmap:

• Marking rejects: The system will directly and visually indicate which panels are “bad,” so operators can remove them immediately without having to search themselves. The operator's role shifts from detector to executor.

Here too, we're taking it step by step: each subsequent phase builds on what the previous one has validated.