D3Former: Dual-Decoder Dual-Transformer Reconstruction Network for Hyperspectral Anomaly Detection

Hyperspectral anomaly detection (HAD) aims to identify and locate anomalous targets in hyperspectral image data that are significantly different from the surrounding background or environment. Most existing deep learning methods for HAD utilize reconstruction errors by reconstructing the background, but they often result in equivalent reconstruction of both background and anomalies, failing to effectively separate them. To address this issue, we propose a novel dual-decoder dual-transformer reconstruction network. Specifically, we introduce a new background feature purification decoding task, which focuses attention on the background direction during image reconstruction. Additionally, we develop an anomaly stripping auxiliary branch that uses autoencoder and dynamic large convolution kernels to mask anomalies and learn local background information, further diminishing the impact of anomalies on the reconstruction process. Furthermore, we design a new cross-guided multi-head self-attention mechanism that integrates multiscale background attributes globally to suppress anomalous pixels and guide the reconstruction of a pure background. Finally, we adopt a synergistic optimization loss to achieve twofold suppression of potential anomalies in both spatial-spectral and frequency domains, resulting in more accurate detection. Extensive experiments on multiple datasets demonstrate that our method outperforms existing state-of-the-art approaches in detection performance.