自然图像对的联合抠图方法

Co-matting of Natural Image Pairs

针对自然图像的抠图问题,研究了一种基于闭形式解的图像对联合抠图方法。根据输入的两张图像与其初始掩像,计算图像对之间的相似矩阵;扩展现有描述单幅图像的抠图拉普拉斯矩阵,提出一种联合抠图拉普拉斯矩阵来刻画图像对;基于联合抠图拉普拉斯矩阵构造能量函数,并通过求解其极小值得到图像对的抠图结果。在多个数据集上进行了实验,实验结果表明新方法具有很好的抠图效果,适用于连续图像序列的抠图。

Aiming at dealing with the problem of natural image matting, this study focused on a closed-form solution to image pairs matting. For given two images and their corresponding trimaps, the affinity matrix between the two images is calculated. The matting Laplacian which characters an image to co-matting Laplacian which characters a pair of images is generalized. The alpha mattes of image pairs were obtained by optimizing the energy function which derived from co-matting Laplacian. To validate the approach, the experiments on several datasets are conducted. The results demonstrate that the method can achieve higher quality and is suitable for matting of a sequence of images.