Overview
This paper introduces a novel two-stage learning approach for solving severely ill-posed inverse problems, particularly in scenarios where training data is scarce. Instead of training a neural network to directly reconstruct an image from raw measurements, the authors propose learning an ensemble of simpler mappings into random, low-dimensional piecewise-constant subspaces (Delaunay meshes). These intermediate projections are then combined to produce a final reconstruction that is highly robust to noise and domain shifts.
The Challenge of Severely Ill-Posed Problems
In fields like geophysics and remote sensing, researchers frequently encounter "inverse problems": attempting to reconstruct an internal structure (such as the Earth's crust) from external measurements (like seismic travel times). These problems are often severely ill-posed, meaning there are far fewer measurements than pixels to reconstruct ($M \ll N$), and the underlying physics makes the reconstruction highly sensitive to noise.
While Deep Learning has revolutionized standard image reconstruction, it typically requires massive "ground truth" datasets. In domains like geophysics, millions of ground truth samples of the Earth's interior simply do not exist. Standard neural networks trained on small or unrelated datasets tend to "hallucinate" features from their training data or become unstable when the noise profile at test time differs from the training environment.
Proposed Method: Regularization by Random Mesh Projections
To overcome these limitations, the authors propose a "divide and conquer" strategy that decomposes the reconstruction into easier sub-tasks.
Stage 1: Decomposition into "Easy" Tasks
Instead of learning the highly complex mapping directly from the data $y$ to the high-resolution image $x$, the method learns projections of $x$ onto random subspaces. The authors utilize Delaunay triangulations to create random piecewise-constant meshes.
Statistical learning theory suggests that the number of samples needed to learn a function scales with its Lipschitz constant (a measure of output sensitivity to input changes). The authors prove that projecting an image into a $K$-dimensional subspace reduces this constant by a factor of $\sqrt{K/N}$. This dimensionality reduction makes the learning task significantly more stable and easier to optimize. They introduce two architectures for this: ProjNet (one network per subspace) and SubNet (a single network handling multiple subspaces simultaneously).
Stage 2: Recomposition
The estimates from these various random meshes are then combined. The authors demonstrate mathematically that, in expectation, this combination process behaves like a deconvolution. By solving a reformulated linear inverse problem using these local estimates, the global structure of the image is accurately recovered.
Key Insights
- Local vs. Global Information: Tomographic measurements (like X-rays or seismic waves) are inherently "non-local"—a single measurement represents an integral along a long path. By projecting into meshes, the networks learn to estimate local averages. These local averages are much more stable and easier to estimate accurately than individual, high-frequency pixel values.
- Dataset Independence: Because the networks learn to estimate local averages rather than memorizing specific object features, the model exhibits remarkable cross-domain generalization. A model trained on a dataset of bridges (LSUN) can successfully reconstruct a seismic cavity. It learns the physics of the projection rather than the statistics of the objects.
Experimental Results
To prove the method does not merely memorize specific image types, the authors trained primarily on the LSUN bridges dataset (20,000 images) and tested on completely different domains, including the BP2004 Geophysics dataset (seismic velocity maps) and a dataset of metal casting X-rays.
- Robustness to Noise: When test-time noise levels differed from training noise, the proposed method significantly outperformed direct inversion. In one evaluation, the proposed method achieved a Signal-to-Noise Ratio (SNR) of 11.56 dB, compared to a direct network's 8.77 dB.
- Handling Missing Data: In "erasure" experiments where measurements were randomly deleted, direct neural networks failed to produce reasonable images. The random mesh approach maintained structural integrity even with significant data loss.
- Qualitative Accuracy: In geophysics tests, the proposed method correctly captured the global shapes of underground cavities, whereas direct inversion often produced sharp but physically incorrect geometries.
Limitations and Future Work
While highly effective, the approach has limitations. Training an ensemble of individual ProjNets is computationally expensive. Although the multi-channel "SubNet" mitigates this, a trade-off between precision and speed remains. Furthermore, the current framework focuses on linear inverse problems; extending this methodology to non-linear physics—where the forward operator $A$ depends on the image $x$ itself—remains an important challenge for future research.
Summary updated