[Paper Review] BSDM: Background Suppression Diffusion Model for Hyperspectral Anomaly Detection

Jitao Ma, Weiying Xie+2 more|arXiv (Cornell University)|Jul 19, 2023

Remote-Sensing Image Classification4 citations

TL;DR

This paper proposes BSDM, a novel background suppression diffusion model for hyperspectral anomaly detection (HAD) that learns latent background distributions without labeled data by using pseudo background noise and a statistical offset module for domain generalization. By innovatively feeding original HSIs into the denoising network during inference, BSDM suppresses complex backgrounds as noise, improving anomaly detection accuracy by up to 26.39% and reducing false alarms, with strong generalization across diverse datasets without retraining.

ABSTRACT

Hyperspectral anomaly detection (HAD) is widely used in Earth observation and deep space exploration. A major challenge for HAD is the complex background of the input hyperspectral images (HSIs), resulting in anomalies confused in the background. On the other hand, the lack of labeled samples for HSIs leads to poor generalization of existing HAD methods. This paper starts the first attempt to study a new and generalizable background learning problem without labeled samples. We present a novel solution BSDM (background suppression diffusion model) for HAD, which can simultaneously learn latent background distributions and generalize to different datasets for suppressing complex background. It is featured in three aspects: (1) For the complex background of HSIs, we design pseudo background noise and learn the potential background distribution in it with a diffusion model (DM). (2) For the generalizability problem, we apply a statistical offset module so that the BSDM adapts to datasets of different domains without labeling samples. (3) For achieving background suppression, we innovatively improve the inference process of DM by feeding the original HSIs into the denoising network, which removes the background as noise. Our work paves a new background suppression way for HAD that can improve HAD performance without the prerequisite of manually labeled data. Assessments and generalization experiments of four HAD methods on several real HSI datasets demonstrate the above three unique properties of the proposed method. The code is available at https://github.com/majitao-xd/BSDM-HAD.

Motivation & Objective

• To address the challenge of complex, confounding backgrounds in hyperspectral images (HSIs) that hinder accurate anomaly detection.
• To develop a generalizable HAD method that does not require manual annotation of background or anomaly samples.
• To improve the discrimination between anomalies and background by learning latent background distributions using a diffusion model.
• To enable effective background suppression during inference without retraining, ensuring low computational cost.

Proposed method

• Proposes a pseudo background noise derived directly from input HSIs, replacing standard Gaussian noise to guide the diffusion model in learning the latent background distribution without any manual labeling.
• Introduces a statistical offset module that embeds statistical features (e.g., mean, variance) of input HSIs into the denoising network to align feature distributions across different domains, enhancing generalization.
• Innovatively modifies the diffusion model inference process by feeding the original HSI into the denoising network, enabling background suppression as noise removal during inference.
• Trains the diffusion model on HSIs to learn to reverse the pseudo background noise, effectively learning to reconstruct clean anomaly-rich images.
• Uses a U-Net-based denoising network with skip connections to preserve spatial-spectral details during the denoising process.
• Employs a noising schedule tailored to the spectral characteristics of HSIs, ensuring stable training and effective background modeling.

Experimental results

Research questions

• RQ1 Can a diffusion model be effectively adapted to suppress complex backgrounds in hyperspectral images without requiring labeled anomaly or background data?
• RQ2 How can a diffusion model be made generalizable across different HSI datasets with varying sensors, spectral characteristics, and scene content?
• RQ3 Can the inference process of a diffusion model be restructured to directly suppress background as noise, rather than generating images?
• RQ4 To what extent does a statistical offset module improve the domain generalization of a background suppression model across diverse HSI datasets?
• RQ5 Does background suppression via BSDM significantly improve the performance of existing HAD methods in terms of detection accuracy and false alarm rate?

Key findings

• BSDM improved the accuracy of 2S-GLRT on the SanDiego dataset by 26.39% while maintaining a low false alarm rate of 0.11%.
• For the DFAE method, BSDM increased accuracy on the Beach dataset from 99.73% to 99.94% and reduced the false alarm rate from 0.03% to 0.01%.
• On the HAD-100 dataset, using Hydice as the training set improved performance compared to using HAD-100 itself, with a 0.77% accuracy gain and a 1.54% reduction in false alarm rate.
• The generalization performance of BSDM dropped by no more than 1.1% in accuracy and 0.8% in false alarm rate when training and test sets were from different domains, demonstrating strong cross-dataset robustness.
• The statistical offset module improved accuracy by 1.45% on the HAD-100 dataset when trained on SanDiego, reducing the performance drop from 1.43% to 1.45% compared to in-domain training.
• BSDM achieved consistent performance gains across four different HAD methods and multiple real-world HSI datasets, confirming its effectiveness and generalization capability.