Hyperspectral anomaly detection via super-resolution reconstruction with an attention mechanism

Hyperspectral anomaly detection aims to classify the anomalous objects in the scene. However, the spatial resolution of the hyperspectral images is relatively low, leading to inaccurate detection of abnormal pixels. Existing methods either ignore the low-resolution problem or leverage super-resolution models to reconstruct the global image to detect abnormal pixels. We claim that reconstructing super-resolution of the global image is unnecessary, while the area where the abnormal target is located should be paid more attention to be reconstructed. In this paper, we propose a super-resolution reconstruction with an attention mechanism for hyperspectral anomaly detection. Our method can automatically extract additional high-frequency information from low-spatial-resolution images and detect abnormal pixels simultaneously. Furthermore, the spatial-channel attention mechanism is adopted to select significant features for reconstructing super-resolution images by assigning different weights to different channels and different spatial-spectral locations. Finally, a regularized join loss function is proposed that balances different tasks by adjusting the relative weight. The experimental results on the public hyperspectral real datasets demonstrate that the proposed method outperforms the state-of-the-art methods.