Two-stage Geometric Information Guided Image Reconstruction

Jing Qin, Weihong Guo

This paper presents GeoCS, a novel two-stage framework for Compressive Sensing (CS) image reconstruction. By combining the directional power of the shearlet transform with an adaptive, weighted Total Variation (TV) approach, the authors successfully reconstruct high-quality images from highly undersampled and noisy data, specifically targeting the "piecewise smooth" textures found in medical and natural imaging.

TL;DR

GeoCS is a two-stage reconstruction method that uses the shearlet transform and adaptive weighted Total Variation to recover complex images from limited measurements. It outperforms standard methods by preserving sharp edges and fine textures while eliminating the "staircase" artifacts typically caused by traditional Total Variation.

What Problem Does This Paper Address?

In Compressive Sensing, the goal is to reconstruct an image from far fewer measurements than the Nyquist-Shannon theorem usually requires. Most existing methods rely on Total Variation (TV) regularization, which assumes images are "piecewise constant" (like a cartoon).

While TV is great for simple shapes, it fails on "piecewise smooth" images—such as medical MRIs or natural scenes—that contain intricate textures and gradients. Standard TV often produces staircase effects, where smooth gradients are incorrectly reconstructed as flat, blocky steps. Furthermore, traditional wavelets struggle to efficiently represent 2D singularities like curved edges. The authors aim to bridge this gap, providing a method that respects both sharp boundaries and smooth transitions.

Methods

The authors propose a two-stage algorithm called GeoCS (Geometric Information Guided CS):

  1. Stage I (Initial Guess): The algorithm solves a standard reconstruction problem using the Shearlet Transform and basic TV. Shearlets are a multi-scale framework specifically designed to be mathematically optimal at representing anisotropic features (edges, corners, and curves) in 2D, making them superior to standard wavelets for image data.
  2. Stage II (Refinement): Using the result from Stage I, the algorithm extracts geometric information (gradients). It uses an edge-stopping function (specifically the Tukey bi-weight function) to create spatially variant weights for the TV term.
  3. In regions with high gradients (edges), the weight is low to preserve sharpness.
  4. In smooth regions, the weight is high to encourage smoothness and suppress noise.
  5. This stage alternates between updating the image and updating the geometric weights in a mutually beneficial loop.

Optimization: To solve these complex minimization problems, the authors utilize the Split Bregman method, which is equivalent to the Alternating Direction Method of Multipliers (ADMM). This allows them to break down the non-differentiable $L_1$ optimization into smaller, faster sub-problems with closed-form solutions.

Datasets

The method was evaluated using simulated partial Fourier sampling (radial trajectories) on three standard test images:

  • Brain MRI: A 512x512 image from BrainWeb, characterized by inhomogeneous contrasts.
  • Barbara: A classic 512x512 test image famous for its complex directional textures (e.g., a tablecloth).
  • Knee MRI: A T1-weighted image showing bone junctions and menisci.

Key Insights and Ideas

  • Anisotropy is Key: The shift from wavelets to shearlets is the "aha moment" for capturing 2D geometry. Because shearlets are directional and elongated, they can represent a curved edge with far fewer coefficients than isotropic wavelets.
  • The "Staircase" Cure: By making the TV penalty spatially variant based on a prior (Stage I), the algorithm avoids the trap of flattening out smooth gradients, which is the primary weakness of traditional CS methods.
  • Continuous vs. Binary Edges: Unlike previous "Edge-Guided" methods that used binary edge masks (edge or no edge), GeoCS uses a continuous weighting function. This allows for a more nuanced reconstruction of gradual intensity changes.

Key Results

The authors compared GeoCS against two state-of-the-art methods: RecPF and EdgeCS.

  • Higher Accuracy: In the Brain MRI test with only 8.79% sampling, GeoCS achieved a Signal-to-Noise Ratio (SNR) of 15.90, significantly higher than RecPF (14.16) and EdgeCS (14.06).
  • Texture Preservation: On the "Barbara" image, GeoCS was uniquely able to recover the fine directional textures of the cloth, whereas other methods produced blurry or over-smoothed results.
  • Noise Robustness: When tested on a noisy Knee MRI ($\sigma=10$), GeoCS maintained a lower Relative Error (0.0952) compared to RecPF (0.1171), proving it is robust to measurement errors.
  • Visual Quality: Qualitatively, the inverted residue maps (difference between ground truth and reconstruction) showed that GeoCS distributes error much more evenly and suppresses artifacts inside complex structures like the skull.

Limitations and Future Work

  • Computational Overhead: While the authors describe the method as efficient, a two-stage iterative process involving shearlet transforms and ADMM is inherently more complex and computationally demanding than simple one-stage TV methods.
  • Initial Guess Dependency: The success of Stage II depends on Stage I providing a "good enough" geometric prior. If the sampling rate is extremely low (e.g., <5%), the initial guess might be too poor to extract meaningful weights.
  • Future Directions: The authors suggest incorporating newer acceleration techniques, such as Nesterov’s accelerated gradient descent, to further speed up the convergence of the Stage II iterations. They also see potential in applying this to other measurement types beyond Fourier samples, such as CT scans.

Summary updated

Abstract

In compressive sensing, it is challenging to reconstruct image of high quality from very few noisy linear projections. Existing methods mostly work well on piecewise constant images but not so well on piecewise smooth images such as natural images, medical images that contain a lot of details. We propose a two-stage method called GeoCS to recover images with rich geometric information from very limited amount of noisy measurements. The method adopts the shearlet transform that is mathematically proven to be optimal in sparsely representing images containing anisotropic features such as edges, corners, spikes etc. It also uses the weighted total variation (TV) sparsity with spatially variant weights to preserve sharp edges but to reduce the staircase effects of TV. Geometric information extracted from the results of stage I serves as an initial prior for stage II which alternates image reconstruction and geometric information update in a mutually beneficial way. GeoCS has been tested on incomplete spectral Fourier samples. It is applicable to other types of measurements as well. Experimental results on various complicated images show that GeoCS is efficient and generates high-quality images.
BibTeX
@misc{qin2014two,
  author = {Jing Qin and Weihong Guo},
  title = {{Two-stage Geometric Information Guided Image Reconstruction}},
  year = {2014},
  eprint = {1409.7450},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1409.7450},
}
Figure