VISION 2026: Validating AI Visual Inspection with Real Defects

Industry analysis | 22 September 2026

For a factory considering AI visual inspection, a good demonstration is only the beginning. The useful question is whether a proposed system can separate the defects that matter when real parts, surface finishes and production conditions change. The forthcoming VISION 2026 programme provides a timely reason to review how inspection data and acceptance trials are prepared.

What the upcoming VISION programme signals

VISION is scheduled for 6–8 October 2026 in Stuttgart. Its first Synthetic Data Symposium, organised with Fraunhofer ITWM on 7–8 October, will examine training-data generation and its practical limits. The organiser identifies scarce and unbalanced industrial image data, together with complex annotation processes as important problems. These are planned sessions, not findings from an event that has already taken place.

In its 16 July 2026 announcement, MVTec describes plans to demonstrate AI-assisted vision development and hardware-accelerated inference. It explicitly calls its AI Vision Solver a prototype. For equipment buyers, the distinction matters: an exhibition roadmap or software demonstration does not establish a production-ready capability in a particular sorting machine.

Separate image generation from acceptance evidence

Synthetic data can be considered during development, but an acceptance trial should be designed around the actual inspection requirement. Define the defect, its permitted limit and the surface on which it must be visible before discussing a training method. A scratch that looks clear in a simulated image may be obscured by real reflections, orientation or contamination.

As a useful technical reference, MVTec AD 2 separates defect-free training and validation images from tests containing normal and anomalous parts. Its test conditions include lighting variations that may be absent from training. This benchmark does not certify any supplier's machine, but it illustrates why repeating a demonstration under one lighting condition provides limited evidence of robustness.

Build a trial around decisions the factory must make

The following is a proposed project workflow, not a published performance result. Keep development samples and final acceptance samples identifiable, and agree which samples may be used to tune the setup. Include acceptable surface variation as well as real rejects, with the quality team resolving borderline examples before the trial.

• Define the decision: record the defect category, relevant drawing limit and required OK, NG or review destination.

• Challenge the views: include realistic part orientations, finish changes and difficult feature locations. Record features that remain outside the agreed optical coverage.

• Report errors separately: count defective parts accepted and acceptable parts rejected by defect category. Include the sample counts and trial conditions instead of presenting one unexplained accuracy percentage.

• Check the complete cycle: review feeding, imaging, decision timing and physical separation at the agreed trial output. Inference time alone does not describe the production line.

MVTec's evaluation workflow provides examples of model evaluation using test data, a confusion matrix and timing information. In a factory proposal, request the equivalent evidence appropriate to the actual software and equipment being offered, rather than requiring or assuming that a particular third-party software package is installed.

Match the inspection task to the machine configuration

A glass-disc setup and a six-side component system present parts differently. The UGV-ER-IV glass-disc sorter requires review of part stability and the camera views for fastener dimensions and visible defects. The MLCC six-side inspection machine addresses a different component-handling and appearance-inspection task. Neither application should inherit a capability merely because it appears in an industry AI announcement.

Prepare the next technical discussion

Send a part drawing, labelled acceptable and defective samples, photographs of the relevant surfaces and your target output. Identify which defects are critical, which variations are acceptable and what evidence the quality team needs for release. Contact Unitecho to review those requirements against an appropriate inspection configuration and a sample-based acceptance procedure.