When non-local similarity meets tensor factorization: A patch-wise method for hyperspectral anomaly detection

Abstract

Hyperspectral anomaly detection (HAD) entails identifying anomaly pixels in hyperspectral images (HSI) that significantly diverge from the background spectral signatures. However, the conventional tensor-based HAD methods typically leverage tensor nuclear norm (TNN) and its variants to exploit the low-rank characteristic on the image-level HSI. Specifically, the omission of non-local similarity features in HAD reduces the potential to leverage pertinent information, leading to the loss of crucial HSI structures and details. In addition, the tensor singular value decomposition (t-SVD) involved in solving the TNN minimization tends to augment the computational time as the data size expands. To address these issues, a novel patch-wise model for HAD that integrates non-local similarity with tensor factorization (NSTF) is proposed. Specifically, instead of using the entire image, we employ aggregated non-local similarity patches to characterize the low-rank structure of HSI, accounting for both spectral correlation and non-local similarity simultaneously and equivalently. Moreover, we introduce tensor factorization to factorize the large-scale tensor generated by grouping non-local similarity patches into two small-scale tensors, which more efficiently exploits its low-rank property of background. Finally, comprehensive experimental results validated on HSI datasets unequivocally establish the superiority of our proposed method over cutting-edge methods.