Lightweight Deep Learning for Real-Time Defect Detection in SMT Component Placement

This study presents a lightweight deep learning approach for real-time defect detection in Surface-Mount Technology (SMT) systems, addressing key challenges in industrial quality control. A dedicated data acquisition system was developed to collect diverse component images directly from the production line, with both offline and online augmentation applied to enhance dataset robustness. Building on this foundation, a modified ResNet-18 architecture was proposed, incorporating Ghost Convolution and Knowledge Distillation to balance accuracy with computational efficiency. Experimental results demonstrate that the optimized model achieves high accuracy (96.5%) while significantly reducing model size and inference latency compared with the baseline ResNet-18. Additional optimization techniques, including quantization and weight pruning, further improved efficiency, with comparisons against MobileNetV2 confirming the competitiveness of the proposed approach. These results highlight the potential of lightweight CNN architectures for SMT component inspection under constrained resources. However, the study remains limited to a specific dataset and experimental setup, and real-world deployment on embedded platforms such as the Raspberry Pi 5 or direct integration into pick-and-place control loops requires further validation.