Deep Image Prior in Frequency Domain for Multispectral Image Demosaicing

Multispectral image demosaicing aims to recover high-quality spectral cubes from mosaic spectral images measured by multispectral filter array imaging technology. Many deep learning-based methods solve this problem by designing supervised end-to-end demosaicing networks. However, the lack of high-quality paired datasets limits the development of this paradigm. To address this issue, this paper presents an unsupervised method inspired by deep image prior. Our method explores deep image prior from the frequency domain perspective. We perform Fourier transform on the target image and constrain both the amplitude component representing brightness distribution and the phase component encoding structural information to achieve tighter feature constraints in the frequency domain. Experiments on multispectral filter array with different pattern sizes demonstrate that our method outperforms other deep image prior-based approaches in qualitative and quantitative evaluations.