PDBSNet: Pixel-Shuffle Downsampling Blind-Spot Reconstruction Network for Hyperspectral Anomaly Detection

Degang Wang, Lina Zhuang, Lianru Gao, Xu Sun, Min Huang, Antonio J. Plaza

Abstract: Recent years have witnessed significant advances of deep learning technology in hyperspectral anomaly detection (HAD). Among these methods, existing unsupervised learning techniques that exploit image reconstruction errors to detect anomalous targets not only generate backgrounds but also reconstruct anomalies to a certain extent, thus failing to recognize anomalies from the original hyperspectral image (HSI). In order to train an efficient background reconstruction network, this article proposes a new pixel-shuffle downsampling blind-spot reconstruction network (called PDBSNet) that operates in a self-supervised fashion. On the one hand, PDBSNet adopts a blind-spot architecture with the receptive field, whose center pixel is set as a blind spot (meaning that the network is unable to see the center pixel), and reconstructs the spectral information of the center pixel using its neighbors. Given that the spectral signatures of anomalous targets are significantly different from those of neighboring pixels, our network tends to generate high reconstruction errors for anomalous pixels and low reconstruction errors for background pixels, and the above characterization of blind-spot architecture is well-suited for the HAD task. On the other hand, we introduce the pixel-shuffle downsampling (PD) technique, which extends the application range of the blind-spot framework due to the fact that it breaks the spatial correlation between pixels in anomalous targets and improves the discrimination between the anomalous pixels and shuffled neighboring backgrounds. Taken together, our PDBSNet integrates the blind-spot architecture and the PD strategy, effectively enhancing the feature representation of the background and weakening the feature expression of anomalies. It is expected to increase the reconstruction error of anomalous objects, while the differences between the original HSI and the reconstructed one are taken as anomaly scores. Extensive experiments on four datasets reveal that PDBSNet is competitive and remarkable concerning other state-of-the-art detectors.