Deep Unfolding Low-Rank Tensor Completion for Large-Scale High-Dimensional Data Recovery

Accurate anomaly detection in monitoring videos from Liquid Rocket Engine (LRE) hot-fire tests critically relies on effective removal of complex interferences, including mist occlusion, distortion, and jitter. Low-Rank Tensor Completion (LRTC) have been widely applied in aeras such as video background modeling, which provides a powerful framework for recovering clean data from high-dimensional, corrupted observations. However, extensive missing entries and severe noise contamination hinder its practical deployment. To address these issues, we propose a scalable and learnable non-convex algorithm termed Deep Unfolded Low-Rank Tensor Completion (DUL-RTC). DULRTC integrates tensor factorization with algorithmic unfolding, achieving low computational complexity. We further develop a hybrid infinite-recurrent unfolding architecture, which generalizes conventional deep unfolding from a fixed number of iterations to a potentially infinite sequence, enabling flexible trade-offs between accuracy and efficiency. Extensive experiments on both synthetic and real-world datasets demonstrate that DULRTC consistently outperforms state-of-the-art baselines. In particular, DULRTC achieves robust tensor recovery under Gaussian noise and random missing patterns, and effectively suppresses complex interferences in monitoring videos from LRE hot-fire tests.