Exploring hyperspectral anomaly detection with human vision: A small target aware detector.

Area of Science:

• Remote Sensing
• Computer Vision
• Image Analysis

Background:

• Hyperspectral anomaly detection (HAD) identifies spectral outliers crucial for applications like environmental monitoring and target recognition where target data is scarce.
• Current HAD methods struggle to replicate the nuanced anomaly detection capabilities of human perception.
• Understanding human visual processing offers a pathway to more robust and intuitive HAD algorithms.

Purpose of the Study:

• To develop a hyperspectral anomaly detection method that mimics human visual perception for improved accuracy.
• To enhance the detection of small or camouflaged targets in hyperspectral images (HSI).
• To create an efficient HAD algorithm suitable for deployment on edge computing devices.

Main Methods:

• Proposed a Small Target Aware Detector (STAD) incorporating saliency maps to capture human-like HSI features.
• Introduced a Small Target Filter (STF) to mitigate the influence of low-confidence regions.
• Developed a knowledge distillation strategy for efficient spectral-spatial feature learning on edge devices.

Main Results:

• STAD demonstrated superior performance in extracting anomalous representations compared to existing methods.
• The STF effectively reduced false positives by filtering low-confidence areas.
• Knowledge distillation enabled a lighter network with preserved detection capabilities for edge deployment.
• Experiments on the HAD100 dataset validated the method's high confidence and excellent performance.

Conclusions:

• The proposed STAD method offers a novel approach to HAD by leveraging human visual perception principles.
• This technique provides a more robust and intuitive solution for detecting anomalies in HSI, especially small or camouflaged targets.
• The efficient design makes HAD applicable to resource-constrained edge devices, expanding its practical utility.