基于图像多尺度分解的前景提取

为了弥补纹理对传统GrabCut提取结果的负面影响,本文分析了图像边缘和颜色分布的尺度特性,结合图像多尺度分解和GrabCut,提出了基于图像多尺度分解的前景提取模型.首先,该模型运用全变分对图像进行多尺度分解得到一系列平滑图像,该分解保护了图像边缘并平滑了纹理,压缩了图像区域颜色的分布范围;其次,将给定平滑图像前景颜色分布表示为高斯混合模型,并运用直方图形状分析方法优化了高斯混合模型的高斯函数个数,弥补了传统固定高斯函数个数的负面影响;最后,根据不同平滑图像的分割结果设计了迭代终止条件,使得从适当的分解尺度中提取前景.与传统前景提取算法相比较,该模型降低了纹理对前景提取的负面影响,其测评分数高于传统算法.

A foreground extraction model on image multiscale decomposition

In order to make up for the negative impact of texture on the extraction results of the traditional GrabCut model, this paper analyzes the scale characteristics of the image edge and color distribution, and combines the image multiscale decomposition and GrabCut to propose a foreground extraction model based on image multi-scale decomposition. This model firstly decomposes an image into a series of smoothed images with the total variation regularization. In this decomposition process, the image edges are preserved, the textures are smoothed, and the color distribution range of the image regions is compressed; secondly, the foreground color distribution of the given smoothed image is represented with the Gaussian mixture model, and the histogram shape analysis method is used to optimize the number of Gaussians in the Gaussian mixture model, which makes up for the negative effects caused by the fixed number of Gaussians; finally, an iterative termination condition is designed according to the segmentation results of different smoothed images, thus the foreground can be extracted from the appropriate decomposition scale. Compared with the traditional foreground extraction algorithm, this model reduces the negative effect of texture on foreground extraction, and the evaluation scores are higher than the traditional algorithms.