基於參考影像的分區影像修補法

Abstract

影像修補技術通常是利用周圍相似的部分來修補一張影像中損毀掉的部分，現今已經有了很多種不同的單張影像修補的技術，然而，單張影像修補的技術無法處理擁有較複雜的紋理的部分，尤其是他們並無法修補出影像中不存在的紋理。 為了解決這個問題，有些論文提出了利用參考影像修補原始影像的方法，但是這些方法在參考影像與原始影像亮度以及色調差異過大的情況下破綻會比較明顯，因此，我們提出了一種基於劃分區域為基礎的影像修補方法，其中分成3個步驟驟:第一，使用SIFT來轉換來源影像的角度，使之與目標影像的角度相同。第二，利用分群法分成不同的區域，在這之前我們會先對影像做一些處理使分群的效果更好，第三，利用分群過後的來源影像做融合，使得來源影像能夠融入目標影像中，並使得其結果看起來盡可能的協調。實驗結果顯示，我們的方法能解決單張影像修補的問題，並且與傳統的影像融合方法比起來，能明顯看出我們方法再影像修補上的優勢。

Image inpainting is widely used to repair damaged regions of a given image. Traditional methods utilize the information from undamaged region in the same image to do the repairing. However, it is difficult to use single image information to repair when its critical parts are missing and the structures of the missing parts cannot be found elsewhere in the image. Therefore, some methods have proposed to use reference images, but these methods have obvious flaws in the situation that reference and target images have distinct color tones and lights. To cope with the problem, we propose a novel region-based image inpainting scheme, which contains 3 major steps. First, we affine the reference image so that its content will have the same viewing angles with the content of the target image. Second, we use K-means to segment the reference image into regions, so that pixels with similar intensities will be grouped together. Third, we proposed a region-based Poisson blending method which, for each damaged region in the target, the corresponding region in the reference image is used for blending and reconstructing the damaged region. Experimental results show the effectiveness and robustness of our proposed algorithm.