一种基于KNN后处理的鲁棒性抠图方法

A ROBUST MATTING METHOD BASED ON KNN POST PROCESSING

目前抠图算法分为采样法和仿射法。采样法求解出的Alpha值往往是不连续的,并且含有很多噪声。对此提出一种基于KNN后处理的鲁棒抠图算法。对未知像素点进行鲁棒性采样,并从中选择较好的样本对作为未知像素的前景和背景样本点,计算出未知像素的初始Alpha值;充分利用KNN方法搜索样本范围宽的优势对初始Alpha进行后处理。实验表明该后处理算法不仅增强了Alpha的视觉效果,同时也提高了Alpha的结果,尤其是实体前景和背景像素结果的准确性。

At present,image matting algorithm is divided into sampling-based method and affinity-based method.The Alpha values obtained by sampling-based method are often discontinuous and contain a lot of noise.To solve the above issues,this paper proposes a robust image matting algorithm based on KNN post-processing.It performed robust sampling of unknown pixels,and the better sample pairs were selected as the foreground and background sample points of unknown pixels,and the initial alpha value of unknown pixels was calculated.The advantages of KNN method were fully utilized to search the sample range and post process the initial Alpha.Experiments show that the post-processing method not only enhances the visual effects,but also improves the accuracy of Alpha results,especially that of opaque foreground and background pixels.