TL;DR
This paper tackles the challenge of reconstructing the locations of point sources in a 2D image when the angles at which the image was projected are unknown. Instead of trying to figure out the angles first, the authors propose a novel pipeline that extracts rotation-invariant geometric features directly from the projection data, which then allows for the reconstruction of the point source locations.
What Problem Does This Paper Address?
Traditional tomography methods often assume that the angles of projection are known. However, in many real-world scenarios, such as in certain imaging techniques or when dealing with historical data, these angles might be unknown or difficult to determine. Reconstructing an image from projections with unknown angles is a significantly harder problem. This paper focuses on a specific case: reconstructing the locations of a sparse set of point sources (like stars in astronomy or particles in microscopy) when the projection angles are unknown and randomly distributed. The core challenge is to extract meaningful geometric information about the point sources without needing to explicitly recover the projection angles.
Methods
The proposed method consists of a three-step pipeline:
- Extracting Rotation-Invariant Features: The authors first compute features from the projection data that are insensitive to the rotation of the underlying point source model. These features are derived from the discrete Fourier transform (DFT) of the projection lines and are shown to be related to Bessel functions. Specifically, they extract two key features: one related to the radial distances of points from the center of mass, and another related to the pairwise distances between points.
- Estimating Distances and Distributions: Two methods are proposed to extract geometric information from these invariant features:
- Prony-based Distance Estimation (PBDE): This method leverages the asymptotic behavior of Bessel functions to directly estimate the radial and pairwise distances using Prony's method.
- Distance Distribution Estimator (DDE): This approach focuses on estimating the distributions of radial and pairwise distances rather than the exact distances themselves. It uses the orthogonality properties of Bessel functions and Hankel transforms.
- Point Source Reconstruction: Finally, the estimated distance distributions (from DDE) are used to reconstruct the locations of the point sources. This is framed as an optimization problem, similar to the unassigned distance geometry problem (uDGP), where the goal is to find a set of point locations whose pairwise distance distribution matches the estimated one. They use a constrained non-convex optimization approach.
Datasets
The paper uses simulated datasets for evaluation. The point source locations are generated randomly in a 2D domain. The projection data is generated by taking projections at uniformly distributed random angles. The authors do not introduce a new benchmark dataset but rather use synthetic data to test their algorithm's performance under various conditions (number of points, signal-to-noise ratio, number of projections).
Key Insights and Ideas
- Bypassing Angle Recovery: The most significant insight is the ability to reconstruct point source locations without explicitly recovering the unknown projection angles. This is achieved by focusing on rotation-invariant geometric features.
- Connection to Distance Geometry: The paper establishes a strong link between unknown view tomography for point sources and the unassigned distance geometry problem (uDGP). This connection allows leveraging existing techniques from distance geometry for reconstruction.
- Bessel Functions as Geometric Signatures: The mathematical derivation shows that the extracted features are directly related to Bessel functions, which encode information about radial and pairwise distances. This provides a theoretical foundation for the method.
- Robustness to Noise: The proposed pipeline demonstrates robustness to additive white Gaussian noise, especially when a sufficient number of projection samples are available.
Key Results
- The proposed pipeline successfully reconstructs point source locations from noisy projection data with unknown angles.
- The PBDE method shows good success rates in recovering radial distances, particularly at higher Signal-to-Noise Ratios (SNR).
- The DDE method, when used for reconstruction, also achieves reasonable success rates, with performance generally improving with a higher number of projection discretizations (M) and higher SNR.
- The method is shown to be robust to noise, with performance degradation being gradual rather than abrupt.
- The paper illustrates the pipeline's effectiveness with visual examples, showing how the estimated distance distributions closely match the true ones, and how the reconstructed point source locations are accurate even from noisy projections.
Limitations and Future Work
- Computational Complexity: While the method is robust, the optimization step for reconstruction can be computationally intensive, especially for a large number of points.
- Approximation Accuracy: The accuracy of the PBDE method relies on the validity of the Bessel function approximations (e.g., for large arguments).
- Future Work: The authors suggest extending this work to more general sparse models beyond point sources, and exploring its application to real-world imaging modalities. Further investigation into improving the efficiency of the reconstruction step could also be a valuable direction.
Summary updated