基於影像深度序列之多層次視訊修補研究

許佳微
To fill the missing area seamlessly from the existing information of a damaged image is a difficult but interesting problem. Recently, many researchers pay their attention to find more effective algorithms. Image inpainting is such a useful technique which can be used to fill the miss area or damaged region automatically. In the thesis, we propose a robust algorithm to repair the missing region of an image, especially for large area.

oposed algorithm does not need to manually adjust the parameters for different images. We define a new filling order to effectively keep the texture structure according to the edge information of a patch. The experimental results demonstrate that our approach can robustly repair the nature images of different textures. Additionally, we present a new framework for video inpainting. The objective of video inpainting is to fill the removed region originally occupied by the selected moving object. Because the selected object may occlude or be occluded by other objects in the video, it is necessary to check which one is closer to the camera when filling the removed region. The contribution of our system is that we can repair the video which contains the removed region of more than two occluded moving objects. The previous works do not address on such issue. In the experimental results, we demonstrate that the objects in a video can be arbitrary removed and we can correctly repair the removed area. Besides, this system can create a novel video which contains the objects coming from others videos.

影像修補技術（image inpainting）主要是運用在修復影像中缺損的區域，使結果部分能逼近原影像的內容，但是傳統的影像修補技術只適用於修復較小的缺損區域， 因此近幾年有許多針對大範圍缺損做修補的相關技術被提出，然而這些方法再修補不同的自然影像時仍有所限制。為克服以上的缺點，本論文提出了一個穩固的影像修補演算法，該演算法主要是著重於大範圍缺損區域做修補。此演算法透過影像中邊的比例（edge ratio）來定義填補順序，其中edge ratio是被用來保留影像的結構資訊。不同與以往的演算法，在復原不同的自然影像時，不需要使用者設定演算法中參數，由實驗結果觀察，不同材質的自然影像皆可有效的透過此演算法來進行復原的動作。另外本論文也提出了一個新的視訊修補（video inpainting）系統，這個系統著重於還原視訊中多個被遮蔽的動態物體；在以往的相關的研究中，大部分提出的系統只用用於還原單一動態物體，主要原因是因為他們缺乏了視訊中的深度資訊。所以我們提出了一個結合深度資訊的視訊修補方法，使本系統可以有效還原視訊中多個被遮蔽的動態物體。從實驗結果可以看出，使用者可以任意的移除視訊中任何物體，而本系統可以正確還原被遮蔽的物體。最後本系統亦透過深度資訊來將合成多個視訊中的場景及物體予以結合，使得合成視訊的內容更加豐富。