• A comprehensive benchmarking of state-of-the-art deep learning models for defect detection in additive manufacturing using thermal imaging is presented. • The study evaluates classical, efficient, hybrid, and custom Convolutional Neural Network (CNN) architectures under consistent training conditions. • The proposed custom CNN achieves the highest accuracy (97.26%) with efficient memory usage (1.91 GB), suitable for deployment in constrained environments. • Thermal imaging effectively identifies key defects such as under-extrusion, over-extrusion, and warping across all models. • Results provide actionable insights for balancing accuracy, computational efficiency, and deployment feasibility in real-time quality control systems. • Systematic benchmark of seven deep learning architectures for thermal imaging-based defect detection in additive manufacturing across five defect categories: under-extrusion, over-extrusion, warping, layer shifting, and normal print. • Proposed custom CNN achieves 97.26% accuracy while ResNet-18 and MobileNetV2 demonstrate competitive performance (97.19% and 97.07%) without thermal-specific modifications. • Comprehensive GPU memory profiling reveals significant variation across architectures (0.84-6.67 GB), with ResNet-18 and Proposed CNN offering optimal balance for resource-constrained deployment. • Narrow performance gap between top architectures (0.19%) indicates that advanced training protocols contribute more to performance than architectural complexity in thermal defect detection applications. Thermal imaging has emerged as a promising approach for real-time defect detection in additive manufacturing, yet the optimal deep learning architecture for this application remains unclear due to limited comparative studies. This study addresses this research gap by conducting a comprehensive benchmarking analysis of seven distinct deep learning architectures for thermal imaging-based defect detection in additive manufacturing. We systematically evaluated classical networks (VGG-16, AlexNet), modern efficient designs (ResNet-18, MobileNetV2, EfficientNet-B0), a hybrid approach (Hybrid VGG-AlexNet), and a proposed custom Convolutional Neural Network (CNN) approach using similar training conditions and a balanced dataset of 21,127 thermal printing images spanning five defect categories: under-extrusion, over-extrusion, layer shifting, warping, and normal prints. Our comparative analysis reveals that multiple architectures achieve competitive performance: top performers reached 97.26% (custom architecture), 97.19% (ResNet-18), and 97.07% (MobileNetV2) accuracy, while hybrid approaches achieved 93.72% accuracy. Also, The Proposed CNN demonstrates a balanced memory profile, with moderate parameter count and efficient memory usage (1.91 GB), making it a promising candidate for deployment in resource-constrained environments. The results demonstrate the system’s effectiveness within controlled experimental conditions and its effectiveness in identifying anomalies such as under-extrusion, over-extrusion, layer shifting and warping. These findings offer practitioners actionable insights for balancing accuracy, computational efficiency, and deployment requirements in additive manufacturing quality control systems.
Shah et al. (Sun,) studied this question.