Hyperspectral image anomaly detection is widely used in agriculture, military, geology, biology, and other fields. To solve the problem of few samples and insufficient utilization of space spectrum characteristics, an anomaly detection algorithm for hyperspectral images combining self-supervision and attention mechanism was proposed. Firstly, a 2D convolutional self-supervised network was used to extract the spectral and spatial features of hyperspectral images to reconstruct the background. The attention mechanism adaptive learning feature channel weights were introduced in the network to optimize the feature. In addition, considering the reduction of image information after background reconstruction, image information entropy was added to the loss function to restrict the feature coding and improve the network performance. Finally, the anomaly value was calculated by the Mahalanobis distance. The proposed algorithm was experimented on two datasets of hyperspectral images from different scenes and compared with seven similar algorithms. The results show that the proposed algorithm has the highest value on area under curve (AUC) .