KDAD: Knowledge Distillation-Based Anomaly Detection for Thermal Infrared Hyperspectral Image

Autoencoder (AE) is extensively utilized in hyperspectral anomaly detection (HAD) tasks owing to its robust feature extraction and image reconstruction capabilities. However, AE lacks constraints on anomaly samples during the training process, leading to the reconstruction part of some anomalies alongside background features, which ultimately diminishes the detection accuracy; additionally, most existing HAD algorithms have been specifically designed for data captured within the visible and near-infrared bands, resulting in a notable gap in methodologies tailored for thermal infrared hyperspectral image (TI_HSI). To address these issues, a knowledge distillation-based anomaly detection (KDAD) model is proposed in this study, aimed at thermal infrared hyperspectral data. KDAD constructs a spatial information map utilizing a dual-window model through the spectral-spatial fusion module, thereby enabling simultaneous fusion of spectral and spatial features via a collaborative stacked AE with dual branches; a residual enhancement module (REM) is introduced based on transfer learning techniques to achieve background purification while forming a distillation AE model comprising an efficient student AE and an intricate teacher AE; meanwhile, REM incorporates a clustering weight generation mechanism that facilitates pixel density-aware category division through dimensionality reduction and clustering processes, and constructs a background-enhanced weight matrix by integrating Mahalanobis distance tensor analysis with dynamic threshold adjustment strategy in order to enforce prior constraints on anomalies; finally, the anomaly detection module formulates an anomaly detection process grounded in clustering techniques and cosine similarity metrics to facilitate high-precision anomaly detection within TI_HSIs. Experimental results on thermal infrared hyperspectral datasets indicate that KDAD markedly enhances background suppression capability and improves anomaly localization accuracy. Furthermore, its detection performance across various scenarios outperforms that of existing algorithms.