김상훈 학생 Multispectral Demosaicking using Bilinear Decomposition with Multichannel Structural Regularization 논문 IEEE Access 게재
2026.05.12 11:05
본 연구실 박사과정인 김상훈 학생이 연구한 “Multispectral Demosaicking using Bilinear Decomposition with Multichannel Structural Regularization” 제목의 논문이 IEEE Access (Impact factor : 3.6, Rank : Q2)에 게재가 확정되었다. 본 연구는 김상훈 학생(1저자)와 이진욱 학생(2저자), 강문기 교수(교신저자)가 진행하였다.
Abstract
In this paper, a multispectral demosaicking algorithm that builds upon bilinear decomposition of a color image is proposed. The proposed algorithm can be summarized into three parts, where the first part constructs a graph-based measure that confines the structure of neighboring pixels. According to the graph, the second part decomposes the raw observation into panchromatic image and chromaticity component. Finally, the third part reconstructs the multispectral data using the two decomposed component. Contrary to previous methods that identify the panchromatic image before chromaticity or spectral difference, the proposed method regularizes the mosaicked chromaticity with bilateral weights that operate as an interpolation stencil to promote similarity between the two, thereby suppressing pattern artifacts. Then, the weights formerly utilized in the identification of the mosaicked chromaticity are applied to interpolate and refine the two components, thus reconstructing chromaticity that has concurrent edges with the panchromatic image. The proposed method demonstrated results comparable to and superior to those of other state-of-theart algorithms in both quantitative and qualitative evaluations, showing less aliasing and pattern artifacts with an accurate visual representation of the trichromatic rendition for the reconstructed multispectral data.
Abstract
In this paper, a multispectral demosaicking algorithm that builds upon bilinear decomposition of a color image is proposed. The proposed algorithm can be summarized into three parts, where the first part constructs a graph-based measure that confines the structure of neighboring pixels. According to the graph, the second part decomposes the raw observation into panchromatic image and chromaticity component. Finally, the third part reconstructs the multispectral data using the two decomposed component. Contrary to previous methods that identify the panchromatic image before chromaticity or spectral difference, the proposed method regularizes the mosaicked chromaticity with bilateral weights that operate as an interpolation stencil to promote similarity between the two, thereby suppressing pattern artifacts. Then, the weights formerly utilized in the identification of the mosaicked chromaticity are applied to interpolate and refine the two components, thus reconstructing chromaticity that has concurrent edges with the panchromatic image. The proposed method demonstrated results comparable to and superior to those of other state-of-theart algorithms in both quantitative and qualitative evaluations, showing less aliasing and pattern artifacts with an accurate visual representation of the trichromatic rendition for the reconstructed multispectral data.
