一种改进的基于KNN颜色线性模型抠图算法

Improved KNN-based Color Line Model Matting Algorithm

介绍了一种减少用户标记和改进的基于KNN(K Nearest Neighbors)颜色线性模型的图像软抠取算法。通过ESCG(Efficient Spectral Clustering on Graphs)算法对输入图像进行谱聚类,用户只需选择某些类中确定的前景、背景像素,便能生成只包含少数未知像素的三分图。基于KNN颜色线性模型的抠图算法将局部平滑假设与非局部原理相结合,但在毛发及前景背景像素近似区域抠取效果并不理想,提出的改进算法将焦点特征添加到特征向量中,最小化基于图拉普拉斯矩阵的二次目标函数并确定未知像素的透明度值。实验表明,改进算法在毛发、孔洞或者图像前景背景近似的区域都能有好的抠取效果。

An improved KNN（K Nearest Neighbors）-based color line model algorithm for efficiently extracting alpha mattes and reducing users＇ marks is presented in this paper. It performs spectral clustering by Efficient Spectral Clustering on Graphs algorithm so that users only need to select pixels which belong to definitely foreground and background, then it enables to generate trimap that only contains small portion of unknown pixels. KNN-based color line model matting algorithm combines color line model and nonlocal principle, it performs poorly when image contains hairs,furs, and similar foreground and background regions. An improved KNN matting algorithm is proposed. The proposed algorithm adds focus informatioh into feature vector, takes advantage of local smoothness and nonlocal principle, then optimization of unknown pixels＇ alpha by minimizing the quadratic object function based on matte Laplacian. The experiments show the improved method performs well in image scenes contain hairs,furs,holes, similar foreground and background regions.