Gaussian-Inspired Attention Mechanism for Hyperspectral Anomaly Detection

Abstract

Hyperspectral anomaly detection (HAD) aims to identify spectrally distinct pixels within a hyperspectral image (HSI). This task necessitates capturing both local spectral information and spatial smoothness, posing a significant challenge for traditional methods. This letter proposes a novel autoencoder framework that leverages a Gaussian-inspired attention mechanism to address this challenge effectively. Specifically, we introduce a novel Gaussian attention layer embedded within the encoder. This layer utilizes a learnable Gaussian kernel to prioritize the local neighborhood of each pixel. This approach effectively captures fine-grained features crucial for background reconstruction. The learned representations are then passed through a deep autoencoder architecture to reconstruct anomaly-free data. Pixels with significant reconstruction errors are subsequently flagged as anomalies. Experiments on several datasets demonstrate the effectiveness of the proposed approach. Compared to existing methods, our framework achieves superior performance in terms of detection accuracy. This finding highlights the potential of Gaussian-inspired attention mechanisms for enhancing HAD. The code is released at: https://github.com/rk-rkk/Gaussian-Inspired-Attention-Mechanism-for-Hyperspectral-Anomaly-Detection .