用於窄波段腸胃道影像之遞迴式解馬賽克演算法研究

Abstract

這篇論文的目標是發展一個專為窄波段腸胃道影像設計的解馬賽克演算法。我們提出了一種遞迴式的演算法，這個方法有效的利用了三原色平面間的相關性以及頻域上高頻頻段的高度關聯性。透過利用疊代的方式，可以提升影像的品質並且消除一些解馬賽克容易產生的失真或是顏色上不正常的變化。然而，使用太多次疊代可能會造成拉練效應並且可能會使整張影像的顏色變得不飽和。為了解決這個問題，我們設立了一個停止疊代的機制，這個機制能判斷這張影像是否適合用疊代來改良品質，適合的話再執行，並且適應性的停止疊代。針對傳統的貝氏彩色濾光陣列，我們和一些已經現存的方法做比較，透過一些常用的測試影像，我們的方法無論在視覺上或是評估指標上都表現得比較好。為了把這樣的演算法適用於窄波段腸胃道影像，彩色濾光陣列裡的綠色成分會被窄波段藍光替換。透過諸多的討論，把貝氏彩色濾光陣列裡的綠色成分替換掉，這樣的排列效果是最好的。針對這個特別設計的彩色濾光陣列，我們的演算法在一個二乘二的彩色濾光區塊中只有三個取樣成分下，還是可以把影像解馬賽克出來。跟傳統的貝氏彩色濾光陣列比較，品質上只損失了1.0185dB。

The objective of this dissertation is to develop a demosaicking algorithm which is suitable for narrow-band gastrointestinal image. We present a hybrid recursive demosaicking method which utilizes the inter-plane and the frequency correlation efficiently. Processing the iteration can improve the image quality and reduce the aliasing and color artifacts. However, too many iterations may cause the zipper effect and tend to de-saturate the color of image. So we set a stop criterion to judge the image is suitable and stop the iteration adaptively. For Bayer CFA, we compare the proposed method using several test images with some existing techniques. Our method performs better both in visual and by the performance measurement indexes. In order to process the demosaicking algorithm for narrow-band gastrointestinal image, the CFA need to withdraw one G component. Through our discussion, we found the CFA which been arranged similar to Bayer pattern can achieve better results. Compare to conventional Bayer CFA, our proposed demosaicking algorithm can lose only 1.0185dB in average by only three samples in 2-by-2 CFA block.