Fast directional image completion

  • Authors:
  • Chih-Wei Fang;Jenn-Jier James Lien

  • Affiliations:
  • Robotics Laboratory, Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan, R.O.C.;Robotics Laboratory, Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan, R.O.C.

  • Venue:
  • PSIVT'07 Proceedings of the 2nd Pacific Rim conference on Advances in image and video technology
  • Year:
  • 2007

Quantified Score

Hi-index 0.00

Visualization

Abstract

We developed a fast image completion system using the multiresolution approach to accelerate the convergence of the system. The down-sampling approach is for the texture-eigenspace training process based on the multi-level background region information. The up-sampling approach is for the image completion process to synthesize the replaced foreground region. To avoid the discontinuous texture structure, we developed the directional and non-directional image completions to reconstruct the global geometric structure and maintain local detailed features of the replaced foreground region in the lowerand higher-resolution levels, respectively. In addition, the Hessian matrix decision value (HMDV) is generated to decide the priority order and direction of the synthesized patch in the replaced region. To avoid the rim effect of the synthesized result, the border of each patch defined as O-shaped pattern is selected for matching comparison instead of using entire patch. Finally, additional texture refinement process is guaranteed to have high-resolution result.