Vision par Ordinateur pour le Contrôle Qualité

Ce qu'elle couvre

Ce programme de niveau praticien forme les équipes qualité et opérations de l'industrie manufacturière à concevoir, construire et maintenir des systèmes d'inspection par vision artificielle. Les participants travaillent sur l'ensemble du pipeline : configuration des caméras et de l'éclairage, workflows d'annotation d'images, entraînement et itération des modèles, puis intégration avec les systèmes MES/SCADA. Le format combine des ateliers pratiques sur des jeux de données réels de défauts et des exercices guidés d'intégration. À l'issue de la formation, les équipes peuvent opérer et ré-entraîner les modèles de façon autonome.

À l'issue, vous saurez

• Design a camera and lighting setup appropriate for a specific defect type and production speed
• Build and manage an image labelling pipeline that produces model-ready annotated datasets
• Train, evaluate, and iteratively improve a defect detection model using real production imagery
• Deploy a trained model to an edge device and connect its outputs to a MES or SCADA rejection workflow
• Define KPIs and monitoring dashboards to track model drift and inspection accuracy over time

Sujets abordés

• Camera selection, lens optics, and structured lighting for industrial inspection
• Image acquisition pipelines and data collection strategies on the production floor
• Annotation tooling and labelling workflows for defect classification and segmentation
• Training object detection and segmentation models (YOLO, Detectron2) on defect datasets
• Model evaluation metrics: precision, recall, false-positive rate in quality contexts
• Iterative retraining and active learning when defect profiles change
• Edge deployment on industrial hardware (NVIDIA Jetson, industrial PCs)
• Integration with MES, SCADA, and PLCs for real-time rejection and alerting

Modalité

Typically delivered as a blended programme over 3-4 weeks: two in-person lab days on-site (or at a partner facility with industrial camera rigs), complemented by guided online modules and weekly coaching calls. Participants bring or are provided with a labelled defect dataset representative of their production context. Hands-on work accounts for roughly 65% of total time. Remote delivery is possible with a hardware kit loan (camera, lighting, edge device) shipped to participants.

Ce qui fait que ça marche

• Involve production operators in labelling sessions — their domain knowledge dramatically improves annotation quality
• Establish a closed-loop feedback process so flagged escapes automatically re-enter the training dataset
• Pilot on a single inspection station with high defect volume before scaling across the line
• Secure OT/IT collaboration early to agree on communication protocols and network segmentation for edge devices

Erreurs fréquentes

• Collecting images under inconsistent lighting conditions, causing models that perform in the lab but fail on the line
• Labelling too few defect samples per class, leading to high false-positive rates that operators learn to ignore
• Treating model deployment as a one-time event rather than building a retraining loop as new defect types emerge
• Skipping MES/SCADA integration design until late in the project, creating costly rework at the OT boundary

Fournisseurs à considérer

• DataScientestdatascientest.com →
• Fraunhofer IPA (Applied AI for Manufacturing)www.ipa.fraunhofer.de/en/expertise/robot-and-assistive-systems/machine-vision.html →
• Coursera – Deep Learning Specialization (Computer Vision track)www.coursera.org/specializations/deep-learning →
• Roboflow (vision-specific training and tooling)roboflow.com →

Sources

• ISO 13374 – Condition monitoring and diagnostics of machines →
• NVIDIA – Vision AI for Industrial Inspection →