具形狀感知的影像縮放技術

Abstract

過去以內容感知為依據的影像縮放技術皆依賴顯著性區域偵測找出人眼視覺上重要的區域，並以此做為維持影像內容的重要度依據。然而，視覺上重要的區域並不代表是最需要被精確維持的。在本篇論文，我們利用影像中重要的輪廓結構來引導整個影像縮放的流程。首先，我們使用一個影像分割技術擷取影像中的重要輪廓結構。接著用一個包含輪廓結構資訊的三角形網格來表示這張影像。我們對這個網格做變形，基於以下幾個條件限制：用來維持曲線形狀的條件，平滑整個網格變形變化的條件，以及一個處理三角形翻轉問題的條件。以上的限制條件可以使用最小平方法快速求解。實驗結果證明我們的方法能產生合理的結果，且能極佳的維持影像內部結構。

Previous content-aware image resizing methods typically account for a saliency map to determine the visually importance of the content and uses the saliency value as a measure of preservation strength. However, it is observed that a visually salient region does not imply the region needs precise preservation. In this thesis, we use the global structure of an image, which is composed of the strong edges of the image, to guide the resizing steps. A powerful image segmentation scheme [1] is first employed to extract the global structure. The image is represented as a triangular mesh that fits the global structure. We deform the mesh with several constraints: curve constraints are used to preserve the shape of the curve, a smoothness constraint is used to smooth the deformation, and a foldover constraint is used to prevent the triangle foldover. The above constraints can be efficiently solved in the linear least-squares sense, and the experiments show that our method can produce convincing results and maintain the overall structure well.