Dependent Hierarchical Beta Process for Image Interpolation and Denoising

Mingyuan Zhou, Hongxia Yang, Guillermo Sapiro, David B. Dunson, Lawrence CarinAISTATS 2011 · OpenReview

Abstract

This paper introduces the Dependent Hierarchical Beta Process (dHBP), a non-parametric Bayesian model designed for sparse image representation. By incorporating spatial covariates into the feature selection process, dHBP allows dictionary atom usage to vary smoothly across an image. This spatial awareness enables the model to achieve state-of-the-art results in image interpolation and denoising, demonstrating particular robustness against extreme pixel loss and non-Gaussian "spiky" noise.

Introduction

In sparse image representation, images are typically decomposed into small patches, each represented as a combination of a few "atoms" from a learned dictionary. Standard models, such as the Beta Process (BP) or the Indian Buffet Process (IBP), often assume these patches are "exchangeable," meaning their spatial location is ignored.

However, natural images possess strong spatial correlations; a patch of a specific texture is likely surrounded by similar patches. Ignoring this context makes models less robust. The dHBP framework addresses this by building a mathematical framework that encourages nearby data points to use similar features while still drawing from a global dictionary.

Methodology

The dHBP builds on the Hierarchical Beta Process by introducing a "local buffet" metaphor. While customers (data patches) pick dishes (features) from a global pool, each customer has a local distribution influenced by the dishes popular in their immediate neighborhood.

Figure 1: The dHBP learns structured dictionary atoms (bottom) and maps their spatial activation (top), showing how feature usage is tied to image location.

Key components of the method include:

  1. Kernel-Based Dependency: A kernel function (such as a Gaussian) defines relationships between patches based on spatial coordinates, influencing feature usage probabilities.
  2. Dictionary Learning: Patches are represented as $x_i = D(s_i \odot z_i) + \epsilon_i$, where $D$ is the dictionary and $z_i$ is a binary mask of active features.
  3. Robustness to Spiky Noise: Unlike models limited to Gaussian noise, dHBP includes a term for "spiky noise" (random bursts of error). Because dHBP expects features to be shared locally, it can distinguish between a legitimate image feature and a spatially isolated noise spike.

Experiments and Results

The model was evaluated on standard grayscale benchmarks, including the Barbara, Boat, and House datasets. These images were subjected to random pixel deletion (up to 80%) and a combination of White Gaussian Noise and spiky noise.

Figure 2: Reconstruction of the 'Boat' image. The model effectively interpolates missing data and removes noise by leveraging spatial context.

Key Findings

  • Interpolation Accuracy: On the "Barbara" image with 80% of pixels missing, dHBP achieved a Peak Signal-to-Noise Ratio (PSNR) of 29.92 dB, significantly outperforming the standard Beta Process (26.90 dB).
  • Denoising Performance: In scenarios with 15% spiky noise, dHBP recovered the "Peppers" image with a PSNR of 29.69 dB, whereas the standard BP failed, reaching only 18.53 dB.
  • Visual Quality: Qualitative analysis shows that dHBP preserves sharp edges and complex textures, such as fine stripes, which standard sparse coding methods often blur.

Discussion and Future Work

The dHBP demonstrates that spatial awareness significantly improves the quality of learned dictionaries. By creating local sub-dictionaries, the model effectively mimics how human vision perceives textures and edges.

However, the current reliance on MCMC inference presents computational challenges, with dHBP taking approximately twice as long per iteration as the simpler BP. Future research will focus on more efficient inference methods, such as Variational Bayes or GPU acceleration, to scale the model for high-resolution video and real-time applications.

Summary updated

BibTeX
@inproceedings{zhou2011dependent,
  author = {Mingyuan Zhou and Hongxia Yang and Guillermo Sapiro and David B. Dunson and Lawrence Carin},
  title = {{Dependent Hierarchical Beta Process for Image Interpolation and Denoising}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2011},
  url = {https://openreview.net/forum?id=EzZMZb-w3uQ},
}
Figure