Abstract:This paper introduces a nonparametric algorithm for bootstrapping a stationary random field and proves certain consistency properties of the algorithm for the case of mixing random fields. The motivation for this paper comes from relating a heuristic texture synthesis algorithm popular in computer vision to general nonparametric bootstrapping of stationary random fields. We give a formal resampling scheme for the heuristic texture algorithm and prove that it produces a consistent estimate of the joint distribution of pixels in a window of certain size under mixing and regularity conditions on the random field. The joint distribution of pixels is the quantity of interest here because theories of human perception of texture suggest that two textures with the same joint distribution of pixel values in a suitably chosen window will appear similar to a human. Thus we provide theoretical justification for an algorithm that has already been very successful in practice, and suggest an explanation for its perceptually good results.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to provide a theoretical basis for a heuristic texture - synthesis algorithm widely used in computer vision and prove its consistency. Specifically, the authors formalize the heuristic texture - synthesis algorithm proposed by Efros and Leung into a framework of non - parametric resampling from stationary random fields and prove that this algorithm can consistently estimate the joint distribution of pixels within a window under the condition of a mixed random field. ### Problem Background Texture synthesis is an important problem in computer vision and computer graphics. The goal is to generate new, seemingly similar but not exactly the same texture images from a given texture sample. This problem is used in computer graphics to "paint" surface textures and in image compression to reconstruct an entire texture from a small sample. ### Heuristic Algorithm The heuristic algorithm proposed by Efros and Leung has achieved very good visual effects in practice, but it lacks theoretical explanations and support from a statistical framework. This algorithm synthesizes new images by directly resampling from the random field of the original image without the need to construct an explicit distribution model. It assumes that the value of a pixel depends only on the values of its neighboring pixels, although it does not explicitly assume that the underlying texture distribution is a Markov random field (MRF). ### Theoretical Contributions To provide a theoretical basis, the authors formalize the algorithm into a non - parametric resampling scheme and prove that under certain conditions, the algorithm can consistently estimate the joint distribution of pixels within a window. Specifically, they prove the following: 1. **Markov Mesh Model (MMM) Version**: For conditionally neighboring regions of a specific shape, the algorithm can consistently estimate the conditional distribution of pixels. 2. **Spiral Resampling Algorithm**: Even when using different pixel - filling orders (such as spiral order), the algorithm can still consistently estimate the joint distribution of pixels within a window. ### Key Assumptions To prove these results, the authors introduce some technical assumptions, including: - **Stationarity and Mixing**: Assume that the random field is strictly stationary and mixing, that is, the dependence between different regions weakens as the distance increases. - **Density Boundedness**: Assume that the marginal distribution and the conditional distribution have bounded density functions. - **Lipschitz Continuity**: Assume that the joint distribution function satisfies certain continuity conditions. - **Kernel Function Properties**: Assume that the kernel function used satisfies some smoothness and symmetry conditions. ### Conclusion Through these theoretical analyses, the authors provide a solid theoretical foundation for Efros and Leung's heuristic texture - synthesis algorithm, explain why the algorithm performs well in practice, and provide theoretical guidance for further improvement and application.

Texture synthesis and nonparametric resampling of random fields

A Constrained Multi-sample Texture Synthesis Algorithm Based on Random Expanding of Circular Patches

Real-time texture synthesis by patch-based sampling

Texture synthesis based on minimum energy cut and its applications

Texture Synthesis And Texture Mixing Using Random Lines

Skeletal Texture Synthesis

A Texture Synthesis Method Based on Adaptive Patches

Fast Texture Synthesis Using Multiple Seeds As Constraints

Reconstructing binary textured images by mean field technique

Asymptotically Admissible Texture Synthesis

Exploring Texture Ensembles by Efficient Markov Chain Monte Carlo-Toward a "trichromacy" Theory of Texture

Salient Structural Elements Based Texture Synthesis

Inverse Texture Synthesis

Filters, Random Fields and Maximum Entropy (FRAME): Towards a Unified Theory for Texture Modeling.

Generating Non-Stationary Textures using Self-Rectification

Cutting and Pasting Irregularly Shaped Patches for Texture Synthesis

Local resampling for patch-based texture synthesis in vector fields

Example-Based Image Synthesis via Randomized Patch-Matching

Exploring the Space of Abstract Textures by Principles and Random Sampling

Order-Independent Texture Synthesis