Texture synthesis using convolutional neural networks with long-range consistency and spectral constraints

Schreiber, Shaun; Geldenhuys, Jaco; Villiers, Hendrik de

doi:10.1109/robomech.2016.7813173

Cited by 8 publications

(2 citation statements)

References 1 publication

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Yet, this algorithm is somehow less generic than the PS algorithm because a different set of features is learned on each texture. From a neurally inspired perspective, convolutional neural networks (CNN) are successfully used to generate textures by simply matching second order statistics of each layer outputs [21]. The synthesized textures show an impressive quality and are often indistinguishable from their original version even when scrutiny is allowed [56].…”

Section: Introductionmentioning

confidence: 99%

“…While the hypothesis that peripheral vision relies on texture statistics generates great progress in understanding the visual cortex, it is still incomplete as it does not account for global scene features such as segmentation [57,43]. Recent deep learning approaches are promising to bridge the gap between low-level texture perception and high-level tasks like scene segmentation and object recognition [21,56,13,52]. Deep neural networks are now used to better probe neural sensitivity in the higher visual cortex [5,40] and their reduced algorithmic load provides more practicable theoretical ground to further model vision [53].…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

The Portilla-Simoncelli Texture Model: towards Understanding the Early Visual Cortex

Vacher¹,

Briand²

2021

Image Processing on Line

View full text Add to dashboard Cite

Texture synthesis is a prolific subarea in computer vision where statistical methods are often successful. The Portilla and Simoncelli (PS) texture algorithm is one of such methods that became very popular and has influenced visual perception studies. For many reasons it can still be considered as a state-of-the art texture synthesis algorithm: (i) it generates textures that are often indistinguishable from the original without scrutiny; (ii) it relies on few parameters compared to recent deep learning methods; (iii) recent algorithms often compare to it. Here, we review the scientific impact of this algorithm and give a detailed explanation. Briefly, the PS algorithm synthesizes a new texture by iteratively imposing to a Gaussian white noise image a set of high-order statistics of wavelet coefficients precomputed on a texture example. After few iterations the initial white noise image is transformed into a texture that is similar to the texture example. We provide a fast C++ implementation, evaluate the effect of the algorithm parameters and illustrate its capabilities with many synthesis examples. In addition, we propose two notable new features to the original implementation: (i) the possibility to interpolate between two textures; (ii) the possibility to handle non-periodicity using the "periodic+smooth" decomposition. Source CodeThe C++ source code and the online demo are accessible at the web page of this article 1 . Compilation and usage instruction are included in the README.txt file of the archive.

show abstract

Section: Introductionmentioning

confidence: 99%