人臉置換下臉部自身陰影及環境光遮擋的修復

Abstract

在本論文中，我們提出了一套在任意光照環境下，模擬出人臉陰影(self-shadow)及環境光遮擋(ambient occlusion)的人臉置換系統。藉由結合陰影圖(shadow map)和重新打光(relighting)的臉部影像，我們重新模擬出類似目標人臉的光影情況。並且使用無接縫影像結合演算法(Drag-and-Drop Pasting)找出最佳接縫邊緣，以維持較好的臉部特徵結果。最後，我們在混合兩張人臉時，採用影像強度梯度的混合(mixing gradients)去保存表情細節，使得最後結果與目標人臉的表情 更加一致也更加生動。

In this thesis, we present a face replacement system, which can simulate the harsh lighting condition such as self-shadow and ambient-occlusion in the target image. First, we use Drag-and-Drop pasting to better maintain the facial saliency with optimal composition boundary. Then we incorporate the mixing gradient method in the blending process to preserve the vivid expression details. In the last step, we propose a method that combines the relighting source image and a shadow map to simulate the self-shadow condition in the target face.