“Efficient Local Texture Regularity Estimation” by Chris and Hofman

  • ©Damkat Chris and Paul M. Hofman

Conference:


Type:


Entry Number: 0019

Title:

    Efficient Local Texture Regularity Estimation

Presenter(s)/Author(s):



Abstract:


    Introduction

    Automatically detecting regular image textures is a challenging task. A regular image texture can be represented by two vectors, a vector pair, that describe the lattice of the texture [Hays et al. 2006]. However, natural images can contain multiple textures that are often only near-regular, i.e. images can be described as being locally regular. To provide such description, the regularity should therefore be estimated at each location. This poster describes an algorithm, Recursive Search Regularity Estimation (RSRE), that efficiently estimates local regularity vectors.

    Regularity estimators in the literature are often based on the autocorrelation function. This has two disadvantages: it is computationally expensive and typically estimates only global regularity. Consequently, it only describes one regular texture per image and cannot describe an image containing multiple (near-)regular textures. An improved method with some similarity to our work is presented by Hays et al. [2006]. There the lattice of near-regular textures is detected by iteratively growing the lattice from an initial seed point driven by local correspondences between texture elements and higher-order lattice geometry. However, also this method is demonstrated only on single textures and it has a high computational complexity. Our method, RSRE, is more efficient and can detect multiple textures. On the other hand, it gives local regularity vectors (see Figure 1) and does not detect the texture lattice.

    RSRE is derived from an efficient, local motion estimator, using the notion that estimation of local regularity and motion is very similar. Both estimates require correspondences between groups of pixels: motion estimation relates pixels between video frames, whereas regularity estimation searches for correspondences within an image or a video frame.

References:


    DE HAAN, G., AND BIEZEN, P. W. A. C. 1994. Sub-pixel motion estimation with 3-d recursive search block-matching. Signal Processing: Image Communication 6, 3, 229–239.

    HAYS, J. H., LEORDEANU, M., EFROS, A. A., AND LIU, Y. 2006. Discovering texture regularity as a higher-order correspondence problem. In 9th European Conference on Computer Vision.


PDF:



Overview Page: