2014/03/31 by Miguel Simões, José M. Bioucas‐Dias, Simões, Miguel +5 · 1 citation
Engineering · #Advanced Image Fusion Techniques #Sparse and Compressive Sensing Techniques #Photoacoustic and Ultrasonic Imaging
paper · pdf · doi:10.48550/arxiv.1403.8098
Hyperspectral remote sensing images (HSIs) are characterized by having a low\nspatial resolution and a high spectral resolution, whereas multispectral images\n(MSIs) are characterized by low spectral and high spatial resolutions. These\ncomplementary characteristics have stimulated active research in the inference\nof images with high spatial and spectral resolutions from HSI-MSI pairs.\n In this paper, we formulate this data fusion problem as the minimization of a\nconvex objective function containing two data-fitting terms and an\nedge-preserving regularizer. The data-fitting terms are quadratic and account\nfor blur, different spatial resolutions, and additive noise; the regularizer, a\nform of vector Total Variation, promotes aligned discontinuities across the\nreconstructed hyperspectral bands.\n The optimization described above is rather hard, owing to its\nnon-diagonalizable linear operators, to the non-quadratic and non-smooth nature\nof the regularizer, and to the very large size of the image to be inferred. We\ntackle these difficulties by tailoring the Split Augmented Lagrangian Shrinkage\nAlgorithm (SALSA)---an instance of the Alternating Direction Method of\nMultipliers (ADMM)---to this optimization problem. By using a convenient\nvariable splitting and by exploiting the fact that HSIs generally "live" in a\nlow-dimensional subspace, we obtain an effective algorithm that yields\nstate-of-the-art results, as illustrated by experiments.\n