內容感知雙邊紋理濾波器

Abstract

一張影像中通常包含了主要的結構(Structures)跟紋理(Textures)，然而會因為紋理的關係沒辦法對這些影像做一些操作，例如細縫裁減(Seam Carving)等，所以需要區分紋理和結構。但以往的方法常常無法有效的區分出比較小的結構和紋理，導致萃取出結構時，較小的結構亦同時遺失。本篇論文改良了Bilateral texture filtering。原本方法是每個像素以一塊Patch來代表，在加上利用Patch Shift來移除影像中的紋理而保留結構，但在邊界處會產生鋸齒以及較小結構會遺失。本論文則是調整patch的大小，在比較重要的結構處的相素使用比較小的patch。我們所提出的方法既快速又簡單，既能移除紋理也能保留細節。

An image often contains structures and textures. Because of textures, we cannot do some operation on an image, such as seam carving. Therefore we need to separate textures and structures. When separating them, we may mix-up small structures and textures and then lose small structures. In this thesis, we improve Bilateral Texture Filtering algorithm. Their method uses a patch to represent a pixel. By patch shift, it can remove textures and preserve structures. However, it could produce some artifacts in edges and lose small structures. Instead of patch shift, we adjust the patch size. In important structures, we use smaller patch to represent a pixel. Our method is fast and easy. It can remove texture and preserve details.