Dictionary learning based sparse representation for hyperspectral anomaly detection

The performance of algorithm is seriously affected by background dictionary in sparse representation based hyperspectral detection. In this paper, a new dictionary learning (DL) sparse representation (SR) based algorithm for anomaly detection in hyperspectral imagery (HSI) was proposed. Principal component analysis (PCA) was used to extract main character of hyperspectral data. Thus, principal component space was established and the feasibility of DLSR in principal component space was proved. Then, the K-SVD algorithm based background dictionary was locally generated using dual window centered at the pixel of interest. Orthogonal matching pursuit (OMP) algorithm and PCA inverse transform were used to reconstruct the original spectrum based on trained dictionary. Thus, regions containing targets were detected by using spectral reconstruction error. Experimental results indicate that the proposed algorithm is effective.