Exemplar-based Texture Synthesis: the Efros-Leung Algorithm
Cecilia Aguerrebere, Yann Gousseau, Guillaume Tartavel
→ BibTeX
    title   = {{Exemplar-based Texture Synthesis: the Efros-Leung Algorithm}},
    author  = {Aguerrebere, Cecilia and Gousseau, Yann and Tartavel, Guillaume},
    journal = {{Image Processing On Line}},
    volume  = {3},
    pages   = {223--241},
    year    = {2013},
    doi     = {10.5201/ipol.2013.59},
% if your bibliography style doesn't support doi fields:
    note    = {\url{https://doi.org/10.5201/ipol.2013.59}}
Cecilia Aguerrebere, Yann Gousseau, and Guillaume Tartavel, Exemplar-based Texture Synthesis: the Efros-Leung Algorithm, Image Processing On Line, 3 (2013), pp. 223–241. https://doi.org/10.5201/ipol.2013.59

Communicated by Julie Digne
Demo edited by Cecilia Aguerrebere


Exemplar-based texture synthesis aims at creating, from an input sample, new texture imagesthat are visually similar to the input, but are not plain copy of it. The Efros–Leung algorithm is one of the most celebrated approaches to this problem. It relies on a Markov assumption andgenerates new textures in a non-parametric way, directly sampling new values from the inputsample.In this paper, we provide a detailed analysis and implementation of this algorithm. The codeclosely follows the algorithm description from the original paper. It also includes a PCA-basedacceleration of the method, yielding results that are generally visually indistinguishable fromthe original results.To the best of our knowledge, this is the first publicly available implementation of thisalgorithm running in acceptable time. Even though numerous improvements have been proposedsince this seminal work, we believe it is of interest to provide an easy way to test the initialapproach from Efros and Leung. In particular, we provide the user with a graphical illustrationof the innovation capacity of the algorithm. Experimentation often shows that the path betweenverbatim copy of the exemplar and garbage growing is somewhat narrow, and that in mostfavorable cases the algorithm produces new texture images by stitching together entire regionsfrom the exemplar.