Defect detection in radiographic weld images is critical for industrial quality assurance. This study proposes a convolutional neural network (CNN) trained on the RIALWEC dataset to automate defect classification. The model achieved a test accuracy of 95.15% with high precision and recall. On the other hand, the confusion matrix indicated low interclass misclassification (<5%), especially among contour resemblance defects. Gradient-weighted Class Activation Mapping was used to highlight the areas within radiographic images that influence the CNN’s defect classification decision, thereby providing an intuitive explanation for its predictions. Cross-dataset validation of GDXRay images demonstrated preliminary generalizability with correct porosity classification despite domain shifts in the imaging protocols. These results highlight the dual capabilities of CNNs, namely, robust diagnostic performance and human-understandable decision logic. The framework enhances the autonomous inspection of welds by providing reliability and accuracy and directly addresses the trust barriers to AI adoption in safety-critical industries, such as aerospace and pipeline construction, where safety is a priority.