Diffusion-Enhanced Hyperspectral Target Detection with Spectral Similarity Purification and Wavelet Cross-Gating

Hyperspectral target detection identifies target pixels in complex backgrounds using a small number of known target priors; this task is complicated by target-sample scarcity, pseudo-sample contamination, and spectral similarity between targets and backgrounds. To address these issues, this paper proposes a diffusion-enhanced hyperspectral target detection method with spectral similarity purification and wavelet cross-gating (SPWCG-HTD). First, a constrained energy minimization detector selects high-confidence pseudo-targets and pseudo-backgrounds from the test image. Global spectral cosine-similarity-weighted reconstruction is then performed within the pseudo-target set to reduce the influence of coarse-detection errors on the generative model. A denoising diffusion probabilistic model subsequently learns the distribution of the purified pseudo-targets, and target-prior-anchored MixUp constructs positive samples. Before the shared Transformer encoder, a single-level Haar wavelet cross-gating module adaptively modulates high-frequency details with low-frequency trends, while a convolutional gated feedforward unit models local inter-band relationships. Finally, the feature space is optimized jointly by a target-aggregation/background-separation loss and a one-positive-one-negative contrastive loss. Experimental results on four real-world scenes demonstrate that the proposed method effectively balances target enhancement and background suppression under limited-sample conditions.