Complex product assembly anomaly inspection based on reference-view synthesis and cross-domain comparison

To address the challenges of high labor intensity, error-proneness, and poor consistency in manual visual inspection of complex aerospace product assembly anomalies, deep learning-based visual anomaly detection methods have been widely adopted. However, existing methods still lack a robust anomaly detection mechanism that effectively adapts to random positional variations, complex industrial environmental interference, and arbitrary viewpoints, while taking the correct assembly state as prior knowledge. To bridge this gap, we propose a reference-view synthesis and cross-domain comparison-based anomaly detection method for complex product assembly. Firstly, a multi-representation reference-domain reconstruction method based on 3D Gaussian splatting is introduced to mitigate the performance degradation of a single representation under complex disturbances. Secondly, a novel view synthesis and cross-domain gap mitigation method is proposed to obtain high-fidelity reference images and minimize the distribution gap between the actual domain and the reference domain. Finally, a multi-type adaptive anomaly detection approach is presented, which achieves anomalous region awareness and adaptive anomaly segmentation through multi-dimensional comparison of cross-domain differences. The experimental results on a real satellite panel anomaly dataset demonstrate the effectiveness and superiority of the proposed method. The proposed method achieves improvements of 5.09%, 9.86%, 9.28%, and 7.65% in precision, recall, F1-score, and mIoU, respectively. Furthermore, we validate that an intelligent vision-assisted inspection scheme can improve the efficiency and accuracy of human-centered anomaly detection for complex aerospace product assembly.