[edit]
Utilization of a Full Convolutional Autoencoder for the Task of Anomaly Detection in Hyperspectral Imagery
Proceedings of 2024 International Conference on Machine Learning and Intelligent Computing, PMLR 245:198-205, 2024.
Abstract
The advancement of artificial intelligence has significantly improved the capability to capture back-ground features in hyperspectral images (HSI), thereby demonstrating commendable performance in the domain of hyperspectral anomaly detection (HAD). The existing approaches, however, still exhibit certain limitations: (1) The deep feature learning process lacks contextual, anomaly constraints, and prior information. (2) The priority reconstruction of the background cannot be ensured by traditional HSI anomaly detectors based on self-supervised deep learning. (3) The utilization of spatial information in hyperspectral images is limited by the fully connected deep network structure of the HSI anomaly detector. The performance of many hyperspectral anomaly detectors is limited by assumptions or presumptions regarding background and anomaly distributions, as these detectors cannot accurately account for the complex real-world distributions. The paper proposes a self-supervised full convolutional autoencoder as a solution to address these issues. The effectiveness and performance of the method were confirmed through evaluation on two real hyperspectral datasets, demonstrating superiority over nine other state-of-the-art methods.