LLE Coordinates for Image Blending
Yufeng Tang, Dongqing Zou, Feng Ding, Jianwei Li, Xiaowu Chen · 2014
Image blending is to paste a source patch onto a target image, which has received considerable attention. While gradient-domain blending is powerful, it suffers from poor scalability. This paper proposes a novel real-time image blending algorithm. We first construct Locally Linear Embedding (LLE) coordinates by utilizing the local geometry of source pixels' color features. Then we propagate the color differences along source and target boundary to all pixels within the source patch. Finally we add color differences to the original color of the source patch, obtaining seamless blending results. To avoid the texture difference between source patch and target image, we further introduce an optimization scheme based on edge-preserving multi-scale decomposition to generate more realistic results. Experiments show that our blending results are visually comparable to gradient-domain methods while possessing real-time performance. Our method can be applied for image stitching and object removal.