使用者導向之圖形結構編修技術

Abstract

目前有許多針對影像中的物件結構進行編修的技術，但已有的技術 多是對基於一些條件限制對物件進行拉伸卻無法使用原有物件生成結 構完全不同的新物件。在本論文中，我們提出一個新技術，藉由使用 者所定義的物件原始結構和期望結構，產生一個新的物件，再具有使 用者所期望的新結構的同時，也具有和原物件相同的風格。我們的系 統藉由沿著新物件的結構資訊，從原始物件中複製影像一個圖塊，並 透過Graphcut的方式將其無縫的放到目標物件上，不斷重複此步驟直 到目標物件生成完成。在此過程中，每一次進行生成的過程都會有許 多可選擇的可能性，我們把此過程變為dynamic programming的方式， 將每個可能性都存起來並且分別進行後續生成的動作直到結束，最後 從中選擇最好的結果，藉以得到和原物件具有相似風格，並且高品質 的新物件。我們的方法只要簡單的讓使用者以單純的線條來描述物件 結構，即可對原始物件進行各種結構上的編修。在本篇論文的最後， 我們將會利用各種物體來展示我們的方法可以如何被應用到影像的編 修上。

Currently there are many techniques for structural image manipulation, but most of them can only stretch an object based on some given constraints, rather than synthesizing an totally structurally-different object whose style is similar to the source object. In this thesis, we proposed a new method which by giving source and target structure information of an object, can synthesize a new object that meets the expected structure of a user, and in the same time, is stylish-similar to the source object. Given the target structure information, our system finds a patch from the source object and puts it on the target object seamlessly by Graphcut, and this step is repeated until the synthesizing of the target object is complete. During the process, each synthesizing step may have many possible source patches to choose. To solve this problem, we consider this process as a dynamic programming that saves the result and cost of each possibility and use each result to continue the process until completion, then we choose the best result among these possibility, thus being able to synthesize a high-quality new object that is stylish-similar to the source object. In our method, a user only needs to input simple strokes to specify the structure of an object to manipulate the structure of an object. We demonstrate many applications to describe how our method can be used in image manipulation.