多模态高斯过程回归抠图  

Gaussian Process Regression Based Multimodal Matting

作　　者：陈秋凤[1] 申群太[1] 刘鹏飞[1,2] 刘建华[3] 

机构地区：[1]中南大学信息科学与工程学院,长沙410083 [2]中国人民解放军95856部队,南京210028 [3]湖南工业大学电气与信息工程学院,株洲412007

出　　处：《小型微型计算机系统》2016年第12期2769-2774,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61473318、60974048、61503131)资助

摘　　要：基于相似度扩散的半监督式抠图是利用图像的局部关系模型构造全局优化函数来求解,现有常用算法的局部关系模型为单一形式,难以适应自然图像的多变特性,提出一种基于高斯过程回归的局部多模态模型.分析自然图像的多变性,据此构建带有常数、线性、非线性3种模态的多模态通用学习框架,通过构造不同的核函数使得局部高斯回归能够适应不同颜色分布情况,后根据高斯回归的置信度来设计模型组合系数,并给出了多模态下的全局优化闭合解.最后从理论上证明两种传统算法为本文算法的特例,从理论上说明了本文算法的先进性.实验表明与传统算法相比,本文算法更能够适应图像的多变特性,在视觉和定量误差上都能取得更好的效果.The affinity propagation based semi-supervised matting derives the global optimized solution from the image local learning model. The local learning models of existing algorithms are too simple to satisfy the diversity of the nature images. Aiming at this question, a local compound model based on Gaussian process regression was prosed. According to the image＇ s diversity, the general multimodal learning framework was built, which included constant, linear and nonlinear models. Through the kernel trick, different Gaussian regression models were carded out to adapt to the different color distribution. Then taking advantage of the confidence of the Gaussian regression,the model combination coefficients were designed and the closed form solution for the multimodal matting was obtained. Finally, the relationship between the proposed algorithm and the traditional algorithm was proved, which indicated the superi- ority of the proposed method. The experiments reveal that the proposed algorithm can better fit the image＇ s diversity, which produce both visually high-quality and quantitatively low-error matting results.

关 键 词：高斯过程回归 局部模型 多模态 半监督式抠图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]