点击次数：
论文类型：会议论文
发表时间：2018-01-01
卷号：10799
页面范围：311-324
关键字：Texture smoothing; Guided bilateral filtering; Structure-adaptive patches
摘要：Almost all of previous works on structure-preserving texture smoothing utilize statistical features of pixels within local rectangular patch to distinguish structures from textures. Since rectangular patches are not aligned to structural boundaries, inexact statistics are inevitable for patches containing both textures and structures. To overcome this problem, a novel structure-preserving texture smoothing approach is proposed via structure-adaptive patches, which conform to local structural boundaries and just contain textures. Specifically, structure adaptive-patches are first generated by several times of classical SLIC superpixel segmentations in the same scale. Secondly, superpixels among different SLIC segmentations are used for computing a guidance image that smooths the fine-scale textures while preserving main structures. Finally, guided bilateral filtering, which incorporates the guidance image into the range filter kernel, is utilized to smooth textures while preserving structural edges. Experimental results demonstrate that the proposed method achieves higher quality results compared to state-of-the-art works.