Provably Convergent Plug-and-play Proximal Block Coordinate Descent Method for Hyperspectral Anomaly Detection

TL;DR

This paper introduces a novel PnP-proximal block coordinate descent method for hyperspectral anomaly detection, leveraging low-rank background modeling, deep learning denoisers, and group sparsity to improve detection accuracy and robustness.

Contribution

It develops a new optimization algorithm integrating deep learning denoisers within a PnP framework for hyperspectral anomaly detection, with proven convergence to stationary points.

Findings

Effective detection of anomalies in hyperspectral images

Outperforms existing methods in noisy scenarios

Robust to Gaussian noise contamination

Abstract

Hyperspectral anomaly detection refers to identifying pixels in the hyperspectral images that have spectral characteristics significantly different from the background. In this paper, we introduce a novel model that represents the background information using a low-rank representation. We integrate an implicit proximal denoiser prior, associated with a deep learning based denoiser, within a plug-and-play (PnP) framework to effectively remove noise from the eigenimages linked to the low-rank representation. Anomalies are characterized using a generalized group sparsity measure, denoted as $\|\cdot\|_{2,\psi}$. To solve the resulting orthogonal constrained nonconvex nonsmooth optimization problem, we develop a PnP-proximal block coordinate descent (PnP-PBCD) method, where the eigenimages are updated using a proximal denoiser within the PnP framework. We prove that any accumulation point of the sequence generated by the PnP-PBCD method is a stationary point. We evaluate the effectiveness of the PnP-PBCD method on hyperspectral anomaly detection in scenarios with and without Gaussian noise contamination. The results demonstrate that the proposed method can effectively detect anomalous objects, outperforming the competing methods that may mistakenly identify noise as anomalies or misidentify the anomalous objects due to noise interference.