关键词: 图像修复, 各向异性, 结构张量, 局部结构测度, 加权, 分形

Abstract:

Aiming at enlarging the searching scope and moderately raising the weights of structure pixels during searching and matching in the process of calculating the best repairing patch for each damaged patch, an image inpainting algorithm based on the weighted fractal under the structure-measurement constraint is presented. First, selected domain blocks are transformed by geometrical transform and isomorphic transform. Using the resulting blocks, the codebook is constructed and serves as the enlarged searching scope. By computing the anisotropic nonlinear structure tensor, the local structure measurement is obtained, in accordance with which the normalized weight map is constructed. The luminance transform is performed, in which we derive its new parameter formulas by minimizing the weighted error between the damaged patch and each domain block in the codebook under the structure-measurement constraint. Lastly, we search in the codebook for the best repairing patch with the minimal weighted error, and then the best repairing patch is used to fill in the damaged patch. Experimental results show that the proposed method can restore damaged structures well and simultaneously make the newly filled region maintain good consistency with the source region. The subjective quality and objective evaluation of the inpainted image are both improved greatly.

Key words: image inpainting, anisotropic, structure tensor, local structure measurement, weighted, fractal