Industrial vision inspection using digital twins: bridging CAD models and realistic scenarios

This study introduces a new industrial visual inspection method that emphasizes the application of computer-aided design (CAD) models. This method significantly reduces the dependence on acquiring and annotating extensive real-scene data, subsequently expediting the development of visual inspection models. The paper highlights two pivotal contributions. Firstly, we introduce a configurable 3D rendering technology that digitally simulates different states of the product, achieving automatic batch generation and labeling of training data. This feature distinguishes our work from existing methods. Secondly, we designed a domain generalization method based on second-order statistics. This approach effectively addresses the domain shift challenge between synthetic and actual production data, enhancing the model’s generalization capabilities. This represents a noteworthy advancement in the field as it boosts the model’s adaptability to real-world scenarios. Our method has demonstrated impressive performance, achieving accuracy rates of 94.30 \(\%\) $\%$ , 96.75 \(\%\) $\%$ , and 97.35 \(\%\) $\%$ on component model classification, motor defect recognition, and rotating motor brush holder datasets, respectively. These results not only validate the efficacy of our domain generalization method but also underscore the potential of using CAD model data for industrial visual inspection. In summary, our research has created a new method for integrating industrial visual inspection into digital twin ecosystems, highlighting the potential for significant improvements in this field.