Traditional time-consuming and error-prone methods of quality control of prefabricated elements can be replaced with state-of-the-art approaches. For example, emerging computer vision and deep learning technologies can be employed in the processes to achieve better performance and quality. To detect defaults and missing welds on prefabricated steel products in real-time, this study enhances the performance, robustness, and extensibility of the computer vision-based method previously proposed by our research team. The design and implementation of a high-performance, accurate, and innovative solution, which corresponds to various manufacturing settings will be investigated in this study.