基于纹理聚类的抠图算法

针对当图像纹理比较丰富时抠图难的问题，基于K近邻(KNN)抠图算法提出了一种纹理聚类抠图（TCM）方法。该方法首先提取出纹理特征;然后用该纹理特征与颜色和位置特征一起构造新的特征空间；接下来在该特征空间上聚类近邻像素以构造Laplacian抠图矩阵；最后利用闭形解求解不透明度。在基准数据集上的实验结果表明，该方法的总排名有显著提升，对于纹理丰富的图像取得了比较好的抠图效果。

Texture clustering matting algorithm

To solve the problem that traditional matting methods do not perform well in highly textured regions, a Texture Clustering Matting (TCM) algorithm based on K-Nearest Neighbor (KNN) matting was proposed. First, the texture features were extracted. Second, a new feature space which contained color, position and texture information was constructed. Third, the matting Laplacian matrix was constructed by clustering neighbors in the new feature space. Last, the opacity was solved by using the closed-form solution. The experiments on benchmark datasets indicate that the overall ranking of the proposed method is significantly improved, which achieves relatively leading matting effect for highly textured image.