應用於數位相機之以小波轉換為主的插補分類器設計
No Thumbnail Available
Date
2008
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
本論文研究中，我們提出一個適用於數位相機色彩插補之小波分類器。其目的是為了解決現行插補演算法中最常使用的插補分類器中頻率響應不一的問題。由於傳統分類器是採用不同階數的微分項所組成，我們的研究中發現在某些頻率下，其頻率響應的強度甚至達到兩倍的差異。這隱含著不同頻率中分類器的組成項是具有不同的權重值，因此，造成方向上的誤判產生錯色的問題。為了解決前述問題，我們提出以小波轉換為基礎的插補分類器來取代傳統的分類器，我們透過小波轉換後得到的高低頻係數矩陣當作其方向判斷的依據。因為在我們的小波分類器中所使用的組成項，均出於小波轉換後的同一個子頻帶。因此，我們解決了頻率響應不同的問題。經由實驗測試，我們所提出的小波分類器的確得到較佳的方向判斷準確性，比傳統分類器平均增加5863點的正確方向判斷點數。應用小波分類器到三個不同的演算法上，PSNR值分別提高了0.50dB、0.19dB與0.20dB，也明顯地大幅改善重建影像視覺上的效果。
In this research, we propose wavelet-based interpolation classifiers for digital still cameras. With them, we solve the issue of the different frequency responses of different terms in traditional interpolation classifiers. The different responses may lead to wrong interpolation directions and result in the color artifacts. To solve this problem, the proposed classifiers are composed from the coefficients in the same subband of wavelet transform domain. Since these coefficients have the identical frequency response, they may lead to more accurate interpolation directions than traditional ones. Simulation results confirm this assumption. The new classifiers averagely increase more 5863 pixels at the correct interpolation directions than traditional classifiers. Applying the proposed classifiers to three demosaicing algorithms, we can elevate peak signal-to-noise ratio up to 0.50dB, 0.19dB and 0.20dB respectively. In addition, the image quality of interpolated images is improved.
In this research, we propose wavelet-based interpolation classifiers for digital still cameras. With them, we solve the issue of the different frequency responses of different terms in traditional interpolation classifiers. The different responses may lead to wrong interpolation directions and result in the color artifacts. To solve this problem, the proposed classifiers are composed from the coefficients in the same subband of wavelet transform domain. Since these coefficients have the identical frequency response, they may lead to more accurate interpolation directions than traditional ones. Simulation results confirm this assumption. The new classifiers averagely increase more 5863 pixels at the correct interpolation directions than traditional classifiers. Applying the proposed classifiers to three demosaicing algorithms, we can elevate peak signal-to-noise ratio up to 0.50dB, 0.19dB and 0.20dB respectively. In addition, the image quality of interpolated images is improved.
Description
Keywords
色彩插補, 解馬賽克, 貝爾圖形, Color interpolation, Demosaicing, Bayer pattern