Deep Densely-Connected Residual Learning for Multispectral Image Demosaicing

Multispectral filter array (MSFA) based multispectral image (MSI) demosaicing is to generate fully defined spectral images at different spectral channels from one snapshot raw mosaicked input. However, it is a very challenging problem, as estimating unknown values from a highly under-sampled input is likely to cause inaccurate artifacts. To address this issue, we propose an end-to-end deep densely-connected residual learning network for snapshot MSI demosaicing. Our proposed method does not require any initial interpolation step, which reduces the computation complexity and computation memory. To avoid vanishing-gradient problem and guarantee efficient information flow for image reconstruction, we apply global residual learning and local residual learning in our proposed architecture. Moreover, stacked densely-connected blocks with skip connections are adopted to make full use of multi-level information extracted from raw mosaicked image. Experimental results show that our proposed model outperforms other state-of-the-art demosaicing methods.