Abstract
This paper presents a novel framework for solving ill-posed inverse problems by exploiting invariant multiscale statistics. Traditional methods often rely on simple geometric priors that fail to capture complex textures, leading to blurry reconstructions. The proposed approach utilizes the scattering transform—a non-linear multiscale representation—to expose strong linear correlations between signal scales. By alternating between measurement consistency and statistical alignment in the scattering domain, the algorithm successfully recovers high-frequency details and sharp textures in tasks such as super-resolution and tomography.
Introduction
In many imaging applications, such as medical tomography or high-resolution sensing, the captured data is often incomplete or noisy. Recovering the original signal from such sparse measurements is a classic "ill-posed" inverse problem. Standard regularization techniques, including Total Variation (TV) or $\ell_1$ minimization, encourage simplicity or sparsity. While effective for piecewise constant signals, these methods often suffer from "regression to the mean," resulting in smoothed or blurry outputs that fail to capture the statistical "vibe" or complex texture of the original data.
The authors address this by focusing on multiscale correlations. They argue that while high-frequency details may be missing in the measurements, their statistical relationship with known low-frequency components can be leveraged to "fill in the blanks" more accurately than traditional convex regularizers.
Methodology
The Scattering Transform
The core of the proposed method is the Scattering Transform, a non-linear representation that uses iterated wavelet transforms followed by a modulus operator. By removing phase information, the transform becomes locally invariant to small translations. Crucially, this process exposes strong linear correlations between different scales that are typically obscured by phase fluctuations in the raw Fourier or wavelet domains. This allows for the linear estimation of missing spectral data within a non-linear space.
Alternating Projection Algorithm
The reconstruction is achieved through an iterative Alternating Projection approach. The algorithm bounces between two primary constraints:
- Measurement Consistency: Ensuring the estimated signal aligns with the physical data captured by the sensor.
- Statistical Consistency: Applying a Linear Minimum Mean Square Error (LMMSE) estimator in the scattering domain to adjust the signal's multiscale statistics to match the target signal class.
Results and Discussion
The framework was evaluated using synthetic models designed to test statistical recovery: the Ising Spin-Glass Model (representing complex, jagged textures) and Cox Point Processes (representing non-homogeneous point-like data).
Key Findings
- Texture Recovery: In a 16x decimation super-resolution task, the proposed method recovered structural details of the Ising model that TV regularization completely blurred.
- Higher-Order Statistics: The authors used Excess Kurtosis as a metric for sharpness and texture. While standard $\ell_1$ methods sometimes achieved lower Mean Squared Error (MSE), the scattering-based approach was significantly superior in recovering the kurtosis of the original signal, producing results that visually and statistically resemble the ground truth.
- Tomography: For limited-angle Radon transforms, the iterative approach matched the cokurtosis tensors of the original signal far better than simple positivity-constrained methods.
Conclusion
The use of invariant multiscale statistics via the scattering transform provides a powerful tool for reconstructing signals with complex textures. By moving beyond simple MSE and geometric priors, this framework captures the underlying statistical structure of signal classes. Future work will focus on reducing the computational cost of the iterative projections and extending the method to natural 3D medical datasets.
Summary updated