Hyperspectral Image Denoising Using Superpixel Segmentation and Graph Laplacian Regularization

Hyperspectral images (HSIs) are often contaminated by complex mixed noise. An efficient mixed noise removal method for HSIs is proposed in this letter, using superpixel segmentation and graph Laplacian regularization (GLR). Using superpixel segmentation, the HSI is divided into superpixel blocks representing homogeneous regions, which possess superior spatial low-rank (LR) properties. In our optimization model, each superpixel block is built as a graph and GLR is used to constrain the LR property of HSI. In addition, the
$\ell _{1}$
norm and a total variation (TV) regularization are included in the optimization model to constrain the sparse noise and preserve the edge information, respectively. Then the augmented Lagrange multiplier (ALM) is used to solve the optimization model. Compared with state-of-the-art HSI denoising methods, the proposed algorithm shows outstanding performance in both simulated and real data experiments.