利用平面結構引導方法移除重複圖案之總結式影像縮放演算法

Abstract

影像縮放技術在數位影像處理上被廣泛的使用。除了傳統利用內插對影像均勻縮放或是手動裁切，許多基於內容感知影像縮放演算法陸續出現，裁切演算法可自動偵測重要區域，保留重要程度較高的區域作為裁切視窗，接縫剪裁演算法則找尋重要程度較少的接縫逐步刪除或複製以達到影像縮放，連續演算法利用扭曲的方式，允許較不重要的區域可以有較大的變形以保留重要區域的大小或比例，多運算子演算法結合不同種運算子，並建立評分機制得到分數最高的輸出影像。以上大多使用低層次資訊為重要程度，例如顏色對比、梯度和熵。 然而上述理論容易因為忽略高層次資訊而導致影像失真，因此就有學者提出了基於高層次資訊的影像縮放演算法，其中之一為重複圖案，重複圖案在真實世界比比皆是，如窗戶或鐵絲網，相關演算法偵測重複圖案，並利用移除重複圖案來縮放影像，但偵測重複圖案時，不論亮度影響或是透視失真都容易造成偵測錯誤。基於此問題，本篇論文提出嶄新的演算法，主要特點有二，第一，一般來說移除重複圖案需要偵測出重複單元的位置或是輪廓，這是個非常容易偵測錯誤的問題，本篇論文不需要偵測出其位置或輪廓，只需計算出重複單元之間的規律性單位即可移除重複圖案。另外，本篇論文提出規律性單位偵測演算法，利用統計的方式偵測出重複單元在矯正仿射空間中的規律間隔，即規律性單位。總結來說，本篇論文提出更加穩定的藉由移除重複圖案縮放影像之演算法，對於大部分具有重複圖案之影像都能得到很好的結果。

Image resizing has been widely used in digital image processing. Besides uniform scaling via interpolation or cropping manually by user, many content-aware based image resizing algorithms have been proposed. Cropping methods detect important regions automatically, and retain the most important region as the cropping window. Seam carving (SC) methods find the least important seam, and remove or duplicate it step by step to resize the image. Continuous methods resize image through warping, preserving the sizes or aspect ratios of important objects by allowing less important regions to be squeezed or stretched. Multi-operator (Multi-Op) methods combine with different operators and use assessment of the targeted image to find the optimal path for the best result. Above methods use low-level-information-based important maps, such as color contrast, gradient, and entropy. However, these methods may result in obvious distortion because of the absence of high-level information. To avoid this problem, some studies take high-level semantics into account. One of them is the regular structure. It exists everywhere in the world, such as windows or barbed wire. These methods detect repeat patterns, and resize images by removing them. However, it is difficult and unstable to detect repeat patterns which are illuminated by different lighting conditions and viewed from different perspectives. To mitigate this problem, we propose a novel algorithm to resize image through removing repeated patterns. In the past, it is necessary to detect certain locations or contours of repeated patterns before removing the repeated pattern. In this thesis, instead of obtaining certain information which is difficult to detect, we just need to find the regularity unit which is easier and more stable to find on the repeated patterns. In addition, we propose a robust and effective algorithm to detect regularity unit in 3D structure by statistics. In summary, we propose a robust algorithm to detect regularity unit and use it to resize an image by removing repeated patterns.