PCM-TV-TFV: A Novel Two-Stage Framework for Image Reconstruction from Fourier Data
Introduction
Reconstructing high-quality images from Fourier measurements—common in MRI, ultrasound, and synthetic aperture radar—remains a significant challenge. These measurements are often non-uniform (jittered) and corrupted by noise, making direct inversion difficult. Traditional reconstruction methods generally fall into two categories: discrete models, which treat the image as a fixed-resolution vector, and continuous models, which treat it as a smooth function.
Furthermore, standard regularization techniques like Total Variation (TV) are effective at preserving sharp edges but often introduce "staircase" artifacts, turning smooth gradients into blocky regions. Conversely, newer Total Fractional Variation (TFV) methods avoid these artifacts but tend to blur sharp boundaries. This paper introduces the PCM-TV-TFV framework, a two-stage approach that bridges continuous and discrete modeling while combining the strengths of TV and TFV regularization.
The PCM Framework
The proposed Projection Correction Modeling (PCM) framework decouples the reconstruction process into two distinct stages: Projection (P-stage) and Correction (C-stage).
- The P-Stage (Projection): This stage addresses the instabilities of non-uniform Fourier data using a continuous model motivated by the multi-scale Galerkin method. By projecting raw measurements onto a continuous subspace spanned by a basis (typically Haar wavelets), the method creates a stable approximation of the image. This acts as a robust foundation for subsequent optimization and alleviates numerical instability caused by non-uniform sampling.
- The C-Stage (Correction): The initial approximation is refined using a novel edge-guided hybrid regularity. The image is partitioned into edges ($\Gamma$) and smooth regions ($\Gamma^c$). TV is applied to the edges to maintain sharpness, while TFV is applied to smooth regions to preserve texture and prevent blockiness. The framework is solved using an Accelerated Proximal Algorithm (based on FISTA) and a Split Bregman scheme.
Edge-Guided Hybrid Regularization
A key innovation is the iterative refinement of the edge mask. The model applies the right mathematical tool to the right part of the image by using a morphologically dilated edge mask. Starting from noisy data, the algorithm recursively updates the set $\Gamma$, allowing it to "discover" the true structure of the image as the reconstruction improves. This ensures that sharp structural boundaries receive TV regularization while the rest of the image benefits from the smoothness of TFV.
Results and Performance
The authors evaluated the framework using synthetic signals, geometric phantoms, and real-world medical imagery (brain MRI and CT scans).
- Quantitative Superiority: In 2D image reconstruction tests with high noise ($\sigma = 0.7$), the PCM-TV-TFV method achieved a Peak Signal-to-Noise Ratio (PSNR) of 24.87, significantly outperforming standard TV (22.60) and Tikhonov regularization (15.51).
- Visual Fidelity: Qualitative results on brain MRIs showed that the proposed method successfully removed the "blocky" artifacts seen in pure TV reconstructions while keeping the edges of brain structures much sharper than pure TFV methods.
- Error Reduction: In surface reconstruction tests, the PCM-TV-TFV model reduced relative error to approximately 32%, compared to 62% for the standard TV-based version of the PCM model.
Conclusion and Future Work
The PCM-TV-TFV framework provides a powerful bridge between continuous and discrete image modeling. By leveraging the complementary strengths of TV and TFV, it offers a robust solution for high-fidelity reconstruction from noisy, non-uniform Fourier data. While the current implementation relies on Haar wavelets, the authors suggest that more sophisticated bases like shearlets or curvelets might yield even better results for images with complex directional textures. Additionally, the framework could be extended to other linear inverse problems, such as the Radon transform used in CT scans.
Summary updated