Abstract: According to the problem of the large visual deviation of matting results due to the less visual information, an unsupervised-matting model based on the visual statistical probability is proposed. The method trains SVM classifier based on the above model to get the SIFT feature point which distinguishes the background area and the higher visual degree foreground target area. And then it generates well-structured Trimap according to the feature points. The unsupervised-matting is achieved by the use of Trimap. Experimental results show that in the case of without user interaction, the model generates [α] mask without large visual deviation, makes good estimate of the foreground object edges and transparency and has better robustness.

Key words: visual statistical probabilistic models, Support Vector Machine（SVM）, natural image matting, unsupervised matting

摘要： 针对无监督抠图因视觉信息较少而存在抠图结果视觉偏差较大的问题，提出一种基于视觉统计概率的无监督抠图模型。该方法根据视觉统计概率模型训练SVM分类器，得到区分背景区域与视觉显著度较高的前景目标区域的SIFT特征点，根据特征点生成结构合理的Trimap，并利用Trimap实现无监督抠图。实验结果表明，在无用户交互的情况下，该模型生成的[α]掩像无较大视觉偏差，对前景目标边缘及透明度做出良好估计并且具有较好的鲁棒性。

关键词: 视觉统计概率模型, 支持向量机（SVM）, 自然图像抠图, 无监督抠图