Abstract
Hyperspectral anomaly detection (HAD) aims to localize pixel points whose spectral features differ from the background. HAD is essential in scenarios of unknown or camouflaged target features, such as water quality monitoring, crop growth monitoring and camouflaged target detection, where prior information of targets is difficult to obtain. Existing HAD methods aim to objectively detect and distinguish background and anomalous spectra, which can be achieved almost effortlessly by human perception. However, the underlying processes of human visual perception are thought to be quite complex. In this paper, we analyze hyperspectral image (HSI) features under human visual perception, and transfer the solution process of HAD to the more robust feature space for the first time. Specifically, we propose a small target aware detector (STAD), which introduces saliency maps to capture HSI features closer to human visual perception. STAD not only extracts more anomalous representations, but also reduces the impact of low-confidence regions through a proposed small target filter (STF). Furthermore, considering the possibility of HAD algorithms being applied to edge devices, we propose a full connected network to convolutional network knowledge distillation strategy. It can learn the spectral and spatial features of the HSI while lightening the network. We train the network on the HAD100 training set and validate the proposed method on the HAD100 test set. Our method provides a new solution space for HAD that is closer to human visual perception with high confidence. Sufficient experiments on real HSI with multiple method comparisons demonstrate the excellent performance and unique potential of the proposed method.
Introduction
Hyperspectral Anomaly Detection (HAD) is a critical task in remote sensing, focusing on identifying pixels that deviate from the statistical or structural properties of the background without prior knowledge of the target. Applications range from environmental monitoring to search-and-rescue operations.
Most existing deep-learning-based HAD methods utilize autoencoders trained on background data, relying on the reconstruction error as an anomaly score. The premise is that the network, optimized for background patterns, will fail to accurately reconstruct anomalous pixels. However, this approach has significant drawbacks: simple anomalies may be easily reconstructed, while complex background regions may yield high errors, leading to false positives. Humans, by contrast, detect anomalies almost effortlessly using visual perception. This paper argues that current algorithms lack this perceptual intuition and proposes a method to bridge the gap between machine processing and human-like saliency.
Proposed Method: Small Target Aware Detector (STAD)
To address the limitations of reconstruction-based methods, we propose the Small Target Aware Detector (STAD), which shifts the detection paradigm to a feature space inspired by human visual perception.
The STAD framework incorporates three key innovations:
- Feature Space Saliency: Instead of raw reconstruction error, STAD utilizes saliency maps. By back-propagating the loss to the input image, the model identifies pixels that significantly influence the network's internal representations, mimicking human visual attention.
- Small Target Filter (STF): Recognizing that anomalies are typically small, the STF combines global Mahalanobis distance with bilateral filtering. This creates a confidence mask that suppresses background noise and focuses the detector on potential small targets.
- Knowledge Distillation: To enable deployment on resource-constrained edge devices, we employ a teacher-student distillation strategy. A complex Transformer-based teacher network, capable of capturing intricate spectral relationships, guides a lightweight CNN student. This allows the student to inherit the teacher's spectral "wisdom" while maintaining the spatial efficiency of a convolutional architecture.
Experimental Results
The proposed method was evaluated on the HAD100 dataset, derived from AVIRIS-NG sensor data. The dataset includes 260 background-only training images and 94 test images with ground-truth labels. To mitigate noise, the authors focused on the first 50 spectral bands.
STAD achieved a state-of-the-art average AUC of 0.9899, outperforming existing methods like AETNet and Auto-AD. Notably, STAD demonstrated superior background suppression ($AUC_{BS}$ of 0.9836) and high robustness, failing on only 2 out of 94 test images.
Furthermore, the distilled student network achieved a throughput of 1.09 Million Pixels per Second (MPS) with a memory footprint of only 2157 MB, making it significantly more viable for real-time edge applications than the teacher network (0.28 MPS, 7445 MB).
Conclusion
STAD introduces a novel approach to HAD by aligning machine detection with human visual saliency. By combining feature-based representation, small target filtering, and efficient knowledge distillation, STAD provides a robust and deployable solution for hyperspectral anomaly detection. Future work will explore more adaptive masking strategies and advanced visualization techniques to further refine detection performance.
Summary updated