Clustering and blending for texture synthesis

Jasvinder Singh, Kristin J. Dana

Research output: Contribution to journalArticlepeer-review

2 Scopus citations


We consider texture modeling techniques which are exemplar-based, i.e. techniques which use several example texture images to learn pixel arrangements. The motivation for these techniques is that the structure of a texture can be characterized by the spatial distribution of pixels in a neighborhood of the image or the multiresolution pyramid. To get another instance of the same type of texture, one can rearrange the pixels as long as certain spatial neighbor relationships are enforced. In this work, we investigate two components of exemplar-based modeling: (1) grouping examples for computational savings during analysis and (2) blending for artifact removal during synthesis. First, we employ clustering in order to group example features and this method provides a significant computational savings without compromising the quality of the texture characterization. Second, we implement techniques for blending to remove border artifacts during the placement stage of texture synthesis. We show that for many textures, the pixel rearrangements can be done without filtering or neighborhood constraints, as long as the blending of borders is done well. Specifically, random rearrangement of texture patches generates a new texture instance of the same texture type when artifacts are removed with blending. We investigate existing blending approaches and introduce a blending method based on image fusion.

Original languageEnglish (US)
Pages (from-to)619-629
Number of pages11
JournalPattern Recognition Letters
Issue number6
StatePublished - Apr 19 2004

All Science Journal Classification (ASJC) codes

  • Software
  • Signal Processing
  • Computer Vision and Pattern Recognition
  • Artificial Intelligence


  • Blending
  • Clustering
  • Texture recognition
  • Texture synthesis


Dive into the research topics of 'Clustering and blending for texture synthesis'. Together they form a unique fingerprint.

Cite this