Bayesian detection of image boundaries

Meng Li, Subhashis GhosalThe Annals of Statistics 2017 · DOI · arXiv

Abstract

Detecting the boundary of an image from noisy observations is a foundational challenge in image processing and segmentation. For $d$-dimensional images, boundaries are often represented as closed, smooth $(d-1)$-dimensional manifolds. This paper presents a nonparametric Bayesian framework utilizing priors indexed by the unit sphere $\mathbb{S}^{d-1}$. We establish optimal posterior contraction rates for Gaussian processes and finite random series priors, specifically employing trigonometric polynomials for 2D images and spherical harmonics for 3D images. For the 2D case, we demonstrate that a rescaled squared exponential Gaussian process on $\mathbb{S}^1$ achieves four critical objectives: guaranteed geometric restriction, minimax optimal adaptation to unknown smoothness, joint inference capabilities, and computational efficiency. Our analysis includes a rigorous study of the reproducing kernel Hilbert space (RKHS) and provides new estimates for modified Bessel functions of the first kind.

1. Introduction

Traditional edge detection methods, such as the Canny operator, typically rely on local derivatives to identify "edge pixels." While effective for simple tasks, these methods often produce fragmented, disconnected boundaries that require significant post-processing to form closed regions suitable for segmentation. Furthermore, local approaches struggle with high noise levels and lack a coherent framework for quantifying global uncertainty or performing joint inference.

This paper shifts the focus from pixel-level classification to global functional estimation. By modeling the boundary as a smooth manifold indexed by a unit sphere, we ensure geometric consistency (e.g., closedness) while leveraging the Bayesian paradigm to provide a full posterior distribution over possible boundaries. This approach allows for the recovery of boundaries even in scenarios with sparse data or significant noise.

2. Methodology

2.1 Boundary Representation and the Star-Shaped Assumption

To simplify the manifold estimation problem, we assume the region of interest is "star-shaped" relative to a known interior point. This allows the boundary to be uniquely represented as a radial function $r(\omega)$ for $\omega \in \mathbb{S}^{d-1}$. In 2D, this reduces the problem to estimating a 1D periodic function on a circle ($\mathbb{S}^1$). In 3D, the boundary is modeled using spherical harmonics.

2.2 Gaussian Process Priors and the SEP Kernel

We employ a Gaussian process (GP) prior for the radial function. Specifically, for 2D images, we utilize the Squared Exponential Periodic (SEP) kernel. To handle unknown smoothness levels ($\alpha$), we introduce a hierarchical "random rescaling" parameter. This allows the model to adaptively determine the appropriate level of regularization without prior knowledge of the boundary's smoothness.

2.3 Theoretical Foundations: RKHS and Bessel Functions

A core contribution of this work is the mathematical analysis of the Reproducing Kernel Hilbert Space (RKHS) associated with the SEP kernel. Unlike standard kernels on Euclidean space, the SEP kernel possesses a discrete spectral measure. We derive new bounds and estimates for modified Bessel functions of the first kind, which characterize the eigenvalues of the GP. These estimates are crucial for proving that the posterior distribution contracts at the minimax optimal rate $n^{-\alpha/(\alpha+1)}$ (up to logarithmic factors).

3. Computational Framework

Efficiency is a primary concern in Bayesian image processing. By exploiting the analytical eigen-decomposition of the SEP kernel using the Fourier basis, we avoid the $O(n^3)$ complexity typically associated with GP matrix inversions. We implement a Gibbs sampler combined with slice sampling that achieves $O(n)$ complexity per iteration, making the approach viable for practical image data.

4. Empirical Results

The proposed method was evaluated through extensive simulations involving various boundary shapes and noise conditions, including Gaussian noise and misspecified models (e.g., mixture normals).

  • Accuracy: In tests with elliptic boundaries, the Bayesian approach significantly outperformed the Maximum Contrast Estimator (MCE), reducing the Lebesgue error (mismatched area) from 6.57 to 0.64.
  • Robustness: The model demonstrated resilience against non-smooth boundaries, such as regular triangles with sharp corners, and remained effective when the noise distribution differed from the model's Gaussian assumptions.
  • Uncertainty Quantification: Unlike frequentist "two-step" methods, the Bayesian framework naturally provides uniform credible bands. This allows researchers to visualize the entire region of uncertainty rather than just a single point estimate.

5. Conclusion and Future Work

This paper provides a robust, theoretically grounded framework for image boundary detection. While the "star-shaped" assumption is a limitation for complex topologies (e.g., objects with holes or multiple disjoint objects), the method offers a powerful tool for applications where boundaries are smooth and noise is prevalent, such as in medical imaging or astronomy. Future research will focus on automating the selection of the interior reference point and extending the framework to handle more complex manifold topologies.

Summary updated

Abstract

Detecting boundary of an image based on noisy observations is a fundamental problem of image processing and image segmentation. For a $d$-dimensional image ($d=2,3,\ldots$), the boundary can often be described by a closed smooth $(d-1)$-dimensional manifold. In this paper, we propose a nonparametric Bayesian approach based on priors indexed by $\mathbb{S}^{d-1}$, the unit sphere in $\mathbb{R}^{d}$. We derive optimal posterior contraction rates for Gaussian processes or finite random series priors using basis functions such as trigonometric polynomials for 2-dimensional images and spherical harmonics for 3-dimensional images. For 2-dimensional images, we show a rescaled squared exponential Gaussian process on $\mathbb{S}^{1}$ achieves four goals of guaranteed geometric restriction, (nearly) minimax optimal rate adapting to the smoothness level, convenience for joint inference and computational efficiency. We conduct an extensive study of its reproducing kernel Hilbert space, which may be of interest by its own and can also be used in other contexts. Several new estimates on modified Bessel functions of the first kind are given. Simulations confirm excellent performance and robustness of the proposed method.
BibTeX
@article{li2017bayesian,
  author = {Meng Li and Subhashis Ghosal},
  title = {{Bayesian detection of image boundaries}},
  journal = {The Annals of Statistics},
  year = {2017},
  doi = {10.1214/16-aos1523},
}
Figure