改进的KNN抠图技术

Improved KNN Matting Method

抠图技术从背景复杂的彩色图像中根据已知像素进行未知像素的掩膜值估计以实现前景的准确提取,是图像处理和影视制作的关键技术之一.通过对KNN抠图的研究分析,针对其在三分图边缘附近颜色相差较小或是图像中前背景复杂交错时抠图质量下降的问题,提出一种增加了纹理特征的改进算法.首先,文中在计算像素的特征向量时,加入纹理特征来增强前景和背景约束,从而特征空间的维数从6维增加至7维;然后由新的归一化的特征向量给出内核函数;最后构造出拉普拉斯矩阵,并使用闭合形式解方法进行优化.实验结果表明,改进后的算法在上述问题上有较好表现,能够得到比原始KNN抠图更好的结果,并给出了几幅有代表性的图像的均方误差值比较进一步加以证实.

Alpha matting is one of the key techniques for image editing and film production,aiming at an accurate extraction of fore- ground elements from a natural image with complicated background by carrying out an alpha estimation of the unknown pixel based on the available information. By studying and analyzing KNN matting algorithm which is belonging to the propagation-based methods, an improved algorithm with a new texture feature attached is proposed in order to deal with the problems that matting quality declines when the color difference at the boundary of trimaps is a bit subtle or the foreground and background layers of matting are complicat- edly mixed. First of all, the texture feature is added to enhance the foreground and background constraints when calculating the pixel feature vectors, and so the dimension of feature space increases from six to seven, furthermore, most of the mathematic formulas and matting properties present new changes;then kernel function is given by the new normalized pixel feature vectors;at last, Laplacian matrix is constructed and a closed-form solution method is adopted to achieve further optimization. The experimental results show that the improved algorithm has good performance in solving the above issues, and can bring about better results than the original KNN matting. The comparison of mean squared error for several representative pictures further confirms the performance of this algorithm.