中文摘要：

针对谱抠图中组件划分整体性不强和前背景组件区分缺乏引导信息的问题,提出了一种带有视觉关注先验的联合特征非监督式抠图方法。首先结合有序径向差分纹理算子,提取图像颜色纹理联合特征,降维并聚类得到初始组件;而后充分利用人类视觉“关注图像中心而忽略边界”这一先验,设计各组件中心先验、边界先验和显著性指标,推导组件的前景和背景概率,并据此概率先验引导最优图层组合得到初步抠图结果;最后结合视觉先验信息和拉普拉斯矩阵的局部特性构造全局优化函数得到优化的抠图结果。实验结果表明,本文算法所提取的前景具有更强的整体性和细节保持能力。

英文摘要：

For spectral matting,the components are not compact enough. It＇s hard to distinguish the foreground components from the background components without heuristic prior. Therefore, this paper pres- ents an unsupervised matting method with visual attention prior. Based on the sorted radial-difference texture operator,the joint color and texture features are abstracted. The features are clustered to obtain the initial components whose dimensions are reduced previously. Taking advantage of the human visual attention prior, the component＇s center prior, boundary prior and saliency are calculated. Then the foreground and background possibilities of the components are calculated. Based on the foreground and background possibilities, a coarse matte with the lowest matting cost is obtained. Finally, the coarse matte is optimized using the visual prior information and laplacian local information. The results reveal that the proposed algorithm has better integrity and detail preserving ability.