使用一階原始對偶演算法進行三原色暨近紅外光影像之去馬賽克

RGB-NIR Demosaicking Using First-Order Primal-Dual Algorithm

杜韋頤(Wei-Yi Duh)

指導教授 ： 盧奕璋

國立臺灣大學/電機資訊學院/電子工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201904355

摘要

去馬賽克（Demosaicking）是相機影像處理流程中常見的步驟，目標是將色彩濾波陣列（Color Filter Array）取樣過的原始資料（Raw Data）內插成全解析度影像（Full-Resolution Image）。紅綠藍去馬賽克是最常見的去馬賽克工作。然而一些時候不僅希望相機可以截取可見光，也希望相機可以截取近紅外光，於是就出現了三原色暨近紅外光去馬賽克（RGB-NIR Demosaicking）的問題，其目標是將含有近紅外光的原始資料內插成四個通道的三原色暨近紅外光全解析度影像。本論文提出以一階原始對偶演算法為核心的三原色暨近紅外光去馬賽克演算法，其特色有三：第一，使用引導影像濾波器作為規律化項，得以充分溝通四個通道的資訊，有效重建影像的高頻成分；第二，將一階原始對偶演算法拆成兩個階段，避免紅綠藍通道與近紅外光通道過度牽制而使結果變差；第三，使用提早終止機制，在預期疊代結果會變差之前將疊代迴圈停下。本論文也採用及改寫了前人的研究，使用色彩校正、高效迴歸先備知識（Efficient Regression Priors）替一階原始對偶演算法提供初始猜測，還有以資料精緻化以及過曝處理作為資料後處理步驟。經過實驗測試之後，本論文演算法可以妥善處理影像的高頻區及物件邊緣，使其不致於出現嚴重的色彩訊噪、暈輪與假影，也能夠有效抹除影像平坦區的雜訊。

關鍵字

三原色暨近紅外光去馬賽克 ； 一階原始對偶演算法 ； 引導影像濾波

並列摘要

Demosaicking is a common process in a camera image signal processing pipeline. It's goal is to interpolate raw data subsampled by the CFA into a full-resolution image. RGB demosaicking is the most common demosaicking task. However, sometimes we want our camera to capture not only visible light but also invisible light. Here comes the problem of RGB-NIR demosaicking, whose goal is to interpolated raw data which contains near-infrared information into a four-channel RGB-NIR full-resolution image. In this work, a RGB-NIR demosaicking algorithm based on first-order primal-dual algorithm is proposed. The algorithm has mainly three features. Firstly, it uses guided image filters as regularization terms to exchange information between four channels, and to effectively reconstruct high frequency components in images. Secondly, it splits first-order primal-dual into two stages, in order not to degrade image quality by overtying RGB and NIR channels. Thirdly, it adopts an early-termination mechanism, which stops iterations when it foresees that the iteration will go wrong. I also adopt and reformulate previous researches, where color correction and efficient regression priors are used to provide better initial guess for first-order primal-dual. And I use data refinement and saturation handling as post-processing. The algorithm can deal with high frequency regions and edges very well, so few chroma noises, halos, and artifacts remain. It can also efficiently remove noise in the flat regions as well.

並列關鍵字

RGB-NIR Demosaicking ； First-Order Primal-Dual Algorithm ； Guided Image Filtering

參考文獻

[1] B. Bayer. Color imaging array. U.S. Patent 3971065, 1976.

[2] M. Brown and S. Süsstrunk. Multi-Spectral SIFT for scene category recognition CVPR 2011, Colorado Springs, CO, USA, pp. 177184, 2011.

[3] A. Chambolle and T. Pock. A first-order primal-dual algorithm for convex problems with applications to imaging. Journal of Mathematical Imaging and Vision, 40(1):120–145, 2011.

[4] Z. Chen, X. Wang, and R. Liang. RGB-NIR multi-spectral camera. Optics Express, vol. 22, no. 5, pp. 4985–4994, 2014.

[5] K. Dabov, A. Foi, V. Katkovnik, K. Egiazarian. Image denoising with block-matching and 3D filtering Proc. SPIE 6064, Image Processing: Algorithms and Systems, Neural Networks, and Machine Learning, 606414, 2006.