Colloquia

Samenvatting

This thesis explores the use of deep learning for automated weld inspection, focusing on the classification of weld defects using combined RGB and depth images. Several convolutional neural network models were developed, each incorporating input modalities differently through unimodal, early fusion, or aggregated output strategies. A comparative study revealed that the fusion models significantly outperformed the unimodal models , with the early fusion model achieving 96.00% accuracy and the aggregated output model reaching 96.86%. These results demonstrate that integrating RGB and depth data enhances classification performance, supporting the reliability of such models for industrial weld inspection.