.png){kind=icon}
.png){kind=icon}
Return
Physics-informed tensor autoencoder with memory for video anomaly detection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.129576.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Video data can be naturally represented as tensors. Despite great progress in anomaly detection with memory-augmented autoencoders, the memory module therein can only handle vectors and inevitably breaks tensor structures, thus leading to performance degradation. Moreover, after the mapping of the encoder, some abnormal features may directly fall into the normal convex polytope, as autoencoders only use the output error to guide the construction of latent variables without imposing any constraint. The memory module can not handle these abnormal features, so that the abnormal observations may not be identified. To solve these problems, we propose a Physics-Informed Tensor AutoEncoder (PITAE) framework, which incorporates both neural networks and physical laws, i.e., tensor operation rules. Specifically, we design a tensor decomposition network followed by an explicit tensor operation to decompose the latent variable into low-rank and sparse components, and only the low-rank component is inputted to the decoder. In this way, we reserve the tensor structure and meanwhile impose a low-rank constraint on the latent variable, thereby compressing the features of normal samples into a ”smaller” region where anomalies are less likely to fall into. Consequently, the non-low-rank anomalies can be identified. But the low-rank anomalies may still not be identified. To further solve this problem, we design a tensor Memory module, and the overall model is named as PITAEM. Finally, based on the proposed framework, we design a novel composite anomaly score to identify anomalies of various kinds. Experiments on various video datasets demonstrate the effectiveness of the proposed method, especially in the small data regime.
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
No organization information available