Benchmark on Inference for Large-Scale Inverse Problems

Welcome to the Benchmark on Inference for Large-Scale Inverse Problems — a BenchOpt benchmark for large scale inverse problems resolution based on DeepInv.

Overview

This benchmark evaluates reconstruction algorithms for large-scale inverse problems across multiple imaging modalities. In inverse problems, we aim to recover the original signal \(x\) from measurements \(y\) following the model:

\[y = Ax + n\]

where \(A\) is the forward operator (e.g., blur, tomographic projection) and \(n\) represents noise.

Datasets

The benchmark includes three imaging scenarios:

  • Tomography (2D/3D): Computed tomography reconstruction from multiple projection operators

  • High-Resolution Color Images: Image restoration from multiple anisotropic Gaussian blur operators

These datasets are multi-operator problems: from a single ground truth, we observe different measurements (e.g., tomography uses different projection angles; natural images use different blur kernels). The goal is to recover the original image from these measurements. The benchmark focuses on large scale images or volumes, of order of magnitude from 1 to 100 million pixels/voxels.

Reconstruction Methods

We leverage the DeepInv library to implement distributed resolution algorithms:

  • Plug-and-Play (PnP): Combines data-fidelity optimization with pretrained denoisers as image priors, offering flexibility and strong performance without task-specific training

Evaluation Conditions

The benchmark assesses solver performance under varying configurations:

  • Image sizes: Testing across different resolution scales

  • Computational resources: From single GPU to multi-GPU distributed setups

Performance is measured through reconstruction quality (PSNR) and computational efficiency (runtime, memory usage).