S3D-Mosaic: Subgradient-Guided Cross-Band Structure Transfer for Spectral Demosaicing

Key points

• Multispectral filter array (MSFA) imaging enables snapshot spectral acquisition, but its per-band sparse sampling causes severe spatial aliasing.1
• Although missing spatial structures can be inferred across bands, wavelength-dependent contrast variations make direct cross-band transfer unreliable.1
• Then we design a compact Directional Sparse Equivariant Difference Convolution (DSE-DConv) for band-level extraction of orientation-selective, high-frequency structure features at low computational cost.1
• Instantiating the framework with DSE-DConv on spectrally adjacent bands yields a cross-band self-structure similarity regularizer for spectral demosaicing, which we optimize via a dual-scan ADMM algorithm that decouples the bidirectional dependency while propagating magnitude-robust structural information across bands.1

Sources (1)

• [1]S3D-Mosaic: Subgradient-Guided Cross-Band Structure Transfer for Spectral Demosaicing
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 17, 03:29 AM

  To address these issues, we propose S3D-Mosaic, a model-based spectral demosaicing method built upon a subgradient-guided cross-band structure transfer framework.

  Multispectral filter array (MSFA) imaging enables snapshot spectral acquisition, but its per-band sparse sampling causes severe spatial aliasing.

Extractive summary: sentences quoted from the sources.