Local complexity difference matting based on weight map and alpha mattes
Local complexity difference matting based on weight map and alpha mattes
复制标题
基于权重图和 alpha 遮罩的局部复杂度差异遮罩
DOI：
10.1007/s11042-022-13223-1
复制
发表时间：
2022-05-23
影响因子：
3.6
通讯作者：
Liang, Yihui
中科院分区：
文献类型：
--
作者：
Feng, Fujian;Huang, Han;Liang, Yihui
图像抠图是计算机视觉中一项重要的图像处理技术，具有广泛的实际应用，包括图像合成、网络广播和电影制作。尽管已经提出了一些方法来提取 alpha matte，但这些方法对于输入图像和真实 alpha matte 之间差异较大的局部复杂区域不够敏感。在本文中，我们设计了一种权重图生成算法，该算法通过测量输入图像和地面实况阿尔法遮罩之间的局部复杂差异来提取每个图像的局部复杂区域。针对像素级损失函数的权重一致性问题，我们提出了一种基于局部复杂度差异的损失函数，可以加强对复杂度差异较大的局部区域的训练。此外，我们在所提出的损失函数和权重图生成算法的基础上设计了一种局部复杂度差异抠图方法，以提高局部复杂度差异图像的抠图精度。为了验证所提出的抠图方法的有效性，在Adobe制作的composition-1k抠图评估数据集上进行了实验。实验结果表明，所提出的权重图生成算法能够有效地提取局部复杂区域。在局部复杂性差异图像的情况下，我们提出的抠图方法优于最先进的抠图方法。
Image matting is an essential image processing technology in computer vision with significant diverse practical applications, including image synthesis, webcasting, and movie production. Although some methods have been proposed to extract the alpha mattes, these methods are not sensitive enough to local complex regions with large differences between the input image and the ground-truth alpha matte. In this paper, we design a weight map generation algorithm, which extracts the local complex region of each image by measuring the local complex differences between the input image and the ground-truth alpha matte. For the weight consistency problem of the pixel-level loss function, we propose a loss function based on local complexity differences, which can strengthen the training on local regions of large complexity differences. Moreover, we design a local complexity difference matting approach on the basis of the presented loss function and weight map generation algorithm to improve the matting accuracy of local complexity difference images. To verify the validity of the proposed matting method, experiments were conducted on the composition-1 k matting evaluation data set produced by Adobe. Experimental results show that the proposed weight map generation algorithm can effectively extract the local complex regions. Our proposed matting method outperforms state-of-the-art matting methods in the cases of locally complexity difference images.