Papers:Texture:Heeger-Bergen-Pyramid
From Dahuawiki
Back to Texture Modeling
[edit]
Pyramid-Based Texture Analysis/Synthesis
- D. J. Heeger and J. R. Bergen.
- SIGGRAPH (1995)
- <Download from ACM portal>
[edit]
Summary
- The paper presents a texture model based steerable pyramid. The basic idea is to decompose a texture image into subbands corresponding to different frequencies and orientations. The steerable pyramid is a combination of Laplacian pyramid and steerable filtering, which first decomposes the image into subbands using standard Laplacian pyramid and then further decomposes each subband into multiple orientation bands with steerable filtering. This process can be inversed for reconstruction.
- The histograms obtained from the decomposed bands respectively are used to characterize the textures. To synthesize a new texture image, it starts from a random noise image, decomposing it into bands, and then use the histogram matching techniques in image processing to reform each subband. The transformed subbands are then combined to form the texture image.
- Instead of working on the original image domain, it manipulates the coefficients of the transformed domain. The author argues that the modification of a coefficient leads to spatially correlated changes.
[edit]
My Comments
- It is a pioneer work in using pyramid-based decomposition in texture modeling. It seems a good idea to use a hierarchial model to respectively capture the spatial features in different scales.
- It uses global histograms of coefficients in all levels independently. It simply ignores the embeded spatial structure and makes it only suitable to describe homogeneous textures.
- The histogram matching is a 1D-mapping. The result is that all coefficients with the same quantized value will be modified to the same value, regardless of where they reside and what their neighborhoods are like. This is clearly another limitation.
In my opinion the major significance of these paper is the introduction of the pyramid-based decomposition paradigm, which has been successfully adopted by a series of later works.
Back to Texture Modeling
