Sparse pixel sampling for appearance edit propagation
Sparse pixel sampling for appearance edit propagation
复制标题
DOI：
10.1007/s00371-015-1094-y
复制
发表时间：
2015-06
期刊：
影响因子：
--
通讯作者：
Tatsuya Yatagawa;Yasushi Yamaguchi
中科院分区：
文献类型：
--
作者：
Tatsuya Yatagawa;Yasushi Yamaguchi
编辑传播是一种使用用户稀疏提供的编辑笔划进行外观编辑的方法。虽然编辑传播有着广泛的应用，但由于需要求解大型线性系统，它在计算上很复杂。为了降低计算成本，基于插值的方法得到了广泛的研究。本研究的灵感来自于一种基于插值的编辑传播方法，该方法使用聚类算法来确定样本。该方法使用插值，它近似编辑参数与样本的凸组合。然而，由于聚类算法生成位于特征空间中的像素集合内部的样本，因此具有凸组合的插值不允许精确重构凸船体外部的像素。为了解决这个问题，本文提出了一种新的近似模型插值图像颜色以及编辑参数使用仿射组合。此外，本文引入稀疏像素采样，同时确定样本的数量和位置以及仿射组合的权重系数。通过更新候选像素来执行稀疏像素采样。不必要的像素被压缩感知丢弃，新的候选像素根据其近似误差被重新采样。本文证明了该模型在图像颜色和编辑参数方面实现了更好的近似，并通过各种实验讨论了该模型的性质。
Edit propagation is an appearance-editing method using sparsely provided edit strokes from users. Although edit propagation has a wide variety of applications, it is computationally complex, owing to the need to solve large linear systems. To reduce the computational cost, interpolation-based approaches have been studied intensely. This study is inspired by an interpolation-based edit-propagation method that uses a clustering algorithm to determine samples. The method uses an interpolant, which approximates edit parameters with convex combinations of the samples. However, because the clustering algorithm generates samples that lie inside the set of pixels in a feature space, an interpolant with convex combinations does not allow for an exact reconstruction of the pixels outside the convex hull. To address this issue, this paper proposes a novel approximation model for interpolating image colors as well as edit parameters using affine combinations. In addition, this paper introduces sparse pixel sampling to determine the quantity and positions of samples and the weight coefficients of the affine combinations simultaneously. Sparse pixel sampling is performed by updating candidate pixels. Unnecessary pixels are discarded with compressive sensing, and new candidate pixels are greedily resampled following their approximation errors. This paper demonstrates that the proposed model achieves better approximation in terms of both image colors and edit parameters, and discusses the properties of the proposed model with various experiments.