基于修正窗口内在变差的结构保持滤波算法

Structure-preserving Filtering via Modified Windowed Intrinsic Variation

针对现有结构保持滤波算法中结构与纹理的分离能力的不足,提出了一种结构似然性度量算子——修正窗口内在变差。该算子通过构造局部加权内在变差,并结合变差修正算子来衡量每个像素点属于边缘结构的似然度;然后将之与双尺度高斯滤波结合构建高质量引导图,并作为联合双边滤波的初始值;通过双边滤波多次迭代获得最终的输出结果。实验结果证实,相较于几种经典算法,本文方法能够在平滑纹理的同时更好地保护图像的精细结构,对边缘结构的处理也更加干净、清晰。

Aiming at the lack of separation ability between structure and texture in the existing structure-preserving filtering algorithms,a new structural likelihood measurement operator,modified windowed intrinsic variation,was proposed.By constructing the locally weighted intrinsic variation and combining with the variation modified operator,the likelihood that each pixel belongs to an edge structure was measured.Then it was combined with a dual-scale Gaussian filtering to construct a high-quality guided image as the initial value of the joint bilateral filtering.The final output result was obtained through multiple iterations of bilateral filtering.Compared with several classic algorithms,the experimental results confirm that this method could preserve the fine structure of the image better while smoothing the texture,and the processing of edge structure is also cleaner and clearer.