Papers:Texture:Bonet-Texture-Synthesis

Multiresolution Sampling Procedure for Analysis and Synthesis of Textured Images

J. S. De Bonet.

SIGGRAPH (1995)

The paper presents a technique to synthesize textures based on multi-scale contrained nonparametric sampling.
The target of the synthesis articulated - the new image is sufficiently different from the original yet still appears as if it was generated by the same underlying stochastic process
The keys of the synthesis method
- Both the input image and the new image are decomposed into subbands in the same way using Laplacian pyramid
- The new texture is synthesizd by sampling successive frequency bands from low resolution to high resolution
- The sampling of each value is conditioned on the corresponding parent struture, which captures the joint occurrence of the features across multiple scales
The contraints are imposed in a thresholded way. As a consequence, the sampling is actually done by nonparametric random rearrangement of the values in input image. The rearrangement are propagated downward along the pyramid through conditioned sampling.

This work is a significant progress in comparison with Heeger and Bergen's. By taking advantages of the cross-scale relation, it succeeds in synthesizing textures with clear spatial structure.
The parent structure here plays an essential role in maintaining the cross-scale consistency, and thus the spatial structure. While, the constrains derived from the augmented filters in each band effectively address the relation between neighboring patches.
The thresholded sampling strategy looks simple, interesting, and efficient. However, it is unclear whether it is comparable to rigorous sampling process in terms of result quality.
The setting of thresholds may notably affect the sampling process.The paper uses an empirical formula to set the thresholds. Can we optimize the thresholds based on some objective?