Progressive self-supervised framework for anomaly detection in hyperspectral images

Hyperspectral image anomaly detection focuses on identifying anomalous targets within hyperspectral images. Many existing techniques rely on explicit priors from hyperspectral datasets and external data, such as sparsity, low rank, and pretrained neural networks. However, these explicit priors can limit the ability to capture subtle complexities in hyperspectral data, and a domain gap problem often exists between hyperspectral and external data. To address these issues, this paper presents a progressive self-supervised framework that eliminates the need for extensive training. The framework learns implicit information from the anomalous hyperspectral image itself, iteratively updating this information to progressively guide the reconstruction of the background hyperspectral image. We introduce an implicit neural prior, termed the progressive prior, through an untrained over-parameterized neural network. This prior enhances the distinction between background and anomaly targets during the iterative process. Since the progressive prior is derived directly from the hyperspectral image data, it does not require external datasets, thereby eliminating the domain gap problem. Extensive qualitative and quantitative evaluations across six hyperspectral image datasets demonstrate that our method achieves an increase of 0.0129 in the area under the curve compared to state-of-the-art anomaly detection methods.