UNet-Based End-to-End Anomaly Detection With Computational Hyperspectral Imaging

UNet-Based End-to-End Anomaly Detection With Computational Hyperspectral Imaging

Most existing hyperspectral anomaly detection methods are developed and validated on numerical datasets, focusing on simulations rather than practical applications. To bridge this gap, we introduce an anomaly detection system that directly utilizes the encoded light intensities (ELIs), avoiding the...

Full description

Saved in:
Bibliographic Details
Main Authors: Weiming Shi, Junren Wen, Yipeng Chen, Yu Shao, Haiqi Gao, Xuehui Wang, Chenying Yang
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Subjects:
Computational hyperspectral imaging
cost matrix
encoded light intensities (ELIs)
end-to-end
UNet
Online Access:https://ieeexplore.ieee.org/document/10988569/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Most existing hyperspectral anomaly detection methods are developed and validated on numerical datasets, focusing on simulations rather than practical applications. To bridge this gap, we introduce an anomaly detection system that directly utilizes the encoded light intensities (ELIs), avoiding the often intricate and time-consuming spectral recovery step in computational hyperspectral imaging. This system employs a UNet-based end-to-end neural network that can adapt to various scene complexities without the need for retraining and further fine-tuning, specifically designed to extract features from the ELI with computational hyperspectral imaging. The UNet is primarily designed for extracting spatial features, and the ELI, which embeds both spatial and implicit spectral information, makes it particularly well suited for leveraging UNet's feature extraction capabilities. Moreover, to optimize the tradeoff between anomaly detectability and background suppression, we incorporate a cost matrix into the loss function to adjust the emphasis on anomalies and the background. The working mode not only simplifies the detection process but also makes it applicable in reality. Our method is evaluated on both generated and experimental datasets, achieving satisfactory performance among various indicators.
ISSN:1939-1404
2151-1535

Similar Items