In a constantly evolving industrial environment, quality control is crucial to ensuring product compliance and optimizing production. This CIFRE thesis is part of a collaboration between TIAMA and LIRIS, UMR 5205 CNRS, aimed at exploring innovative solutions to improve the performance of quality control systems. TIAMA, which specializes in the automatic inspection of hollow glass products, has developed control systems that use industrial vision and multimodal fusion technologies to detect manufacturing defects.
However, current systems rely on predefined models that do not sufficiently integrate continuous learning and dynamic adaptation to production variations and customer needs. Material variability, changes in lighting conditions, and developments in industrial processes mean that these systems can be improved. This thesis aims to develop adaptive and self-supervised learning models that allow the model to adjust dynamically without heavy human intervention.