针对视频序列分类的快速抠像技术

针对视频序列分类的问题提出了一种快速抠像技术.根据视频序列间的相关度进行关键帧的区分,得到关键帧、序列间变化细微的非关键帧、序列间变化较大的非关键帧;对于关键帧,采用闭合式的抠像方法来进行处理,获得透明度值、前景像素值和背景像素值;对于变化细微的非关键帧,提出了一种基于帧间连续性的透明度值估计和优化方法;对于变化较大的非关键帧,提出了一种基于特征流传递的机制来传递关键帧的有效信息.实验结果表明,最终在获得与传统方法相比可接受的抠像效果条件下,这种快速抠像技术缩短了处理时间.

Rapid Image Matting Method Based on Categorized Video Sequences

This paper proposes a rapid image matting method based on categorized video sequences. The image sequences in a video are classified into three categories:key frames, trivial-changed frames and variant-changed frames. Different image matting algorithms are adopted to treat each image sequences category. First, the method uses closed-from image matting algorithm to process the key frames and get the alphas, foreground colors and background colors. Then, the proposed method uses the known values of key frames to estimate and optimize the alphas of the trivial-frames based on frame-coherence. After that, the matting algorithm is processed by utilizing a SIFT Flow based information propagation strategy to transfer informs from key frames to variant-changed frames. The proposed method achieves acceptable matting results and more higher efficiency compares to traditional methods.