Deep Learning Driven 2D/3D Image Analysis Techniques for SurfaceCharacterisation and Defect Detection in Additive Manufacturing: AnOverview

Introduction: Deep learning has rapidly transformed Additive Manufacturing (AM) quality control by enabling advanced 2D and 3D image analysis for defect detection and surface characterisation. Current approaches utilising YOLO, U-Net, and 3D CNNs have demonstrated significant improvements. However, critical research gaps remain, particularly in data scarcity, cross-domain generalisation, and integration with legacy manufacturing systems. This review bridges these gaps by comprehensively synthesising recent advances (2020-2025) and highlighting emerging solutions through multi-modal sensor fusion and physics-informed neural networks. Methods: This systematic review analysed over 200 peer-reviewed publications from major scientific databases (Scopus, IEEE Xplore, Elsevier, SpringerLink). Selection criteria focused on empirical AM studies with quantitative performance metrics. Data extraction encompassed imaging modalities, neural architectures, datasets (VISION, EOSTATE PowderBed with over 1.2 million images, NEU-CLS, Defect Spectrum), evaluation protocols (IoU, mAP, F1 score), and real-time deployment strategies. Results: State-of-the-art 2D detection models (GDCP-YOLO, YOLOv8-enhanced) achieved a mean average precision (mAP@0.5) of 97.5% with real-time speeds of 71.9 FPS. Volumetric 3D models (ResidualUNet3D+SE, THz 3D-CNN) achieved subsurface defect detection accuracies of 98.9% with IoU values exceeding 88.4%. Edge AI systems (NVIDIA Jetson Xavier) demonstrated inference times of 118.83 ms, enabling closed-loop control, reducing defect volumes from 10.7% to 1.3%. Multi-modal sensor fusion achieved 98.5% defect prediction accuracy without thermal modalities. Discussion: Despite remarkable technical progress, challenges remain, such as (1) data scarcity, only 15 public datasets identified despite significant research activity, (2) cross-domain generalization failure models show 16.9-22.5% performance degradation across materials/systems, (3) interpretability gaps: black-box predictions limit adoption in safety-critical aerospace/ biomedical applications, and (4) integration barriers retrofitting costs prohibitive for existing equipment. Recent advances in self-supervised learning (70% annotation reduction), domain adaptation techniques, and explainable AI (XAI) with LIME/SHAP show promise but require standardised validation frameworks. Conclusion: This review demonstrates that deep learning-driven AM quality control has matured substantially, with real-time capabilities and high detection accuracies now feasible. However, transitioning from laboratory prototypes to production deployment requires addressing data availability, model generalisation, and regulatory compliance. Future directions emphasise physics-informed neural networks, federated learning for cross-organisational data sharing, and uncertainty quantification methods essential for Industry 4.0 manufacturing paradigms, enabling zero-defect production.