BS3LNet: A New Blind-Spot Self-Supervised Learning Network for Hyperspectral Anomaly Detection

Abstract

Recent years have witnessed the flourishing of deep learning-based methods in hyperspectral anomaly detection (HAD). However, the lack of available supervision information persists throughout. Additionally, existing unsupervised/semi-supervised learning methods to detect anomalies utilizing reconstruction errors not only generate backgrounds but also reconstruct anomalies to some extent, complicating the identification of anomalies in the original hyperspectral image (HSI). In order to train a network able to reconstruct only background pixels (instead of anomalous pixels), in this paper we propose a new blind-spot self-supervised learning framework (called BS <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> LNet) that generates training patch pairs with blind-spots from a single HSI and trains the network in self-supervised fashion. The BS <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> LNet tends to generate high reconstruction errors for anomalous pixels and low reconstruction errors for background pixels due to the fact that it adopts a blind-spot architecture, i.e., the receptive field of each pixel excludes the pixel itself and the network reconstructs each pixel using its neighbors. The above characterization suits the HAD task well, considering the fact that spectral signatures of anomalous targets are significantly different from those of neighboring pixels. Our network can be considered a superb background generator, which effectively enhances the semantic feature representation of the background distribution and weakens the feature expression for anomalies. Meanwhile, the differences between the original HSI and the background reconstructed by our network are used to measure the degree of anomaly of each pixel, so that anomalous pixels can be effectively separated from the background. Extensive experiments on two synthetic and three real datasets reveal that our BS <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> LNet is competitive with regards to other state-of-the-art approaches.