Multi-frame Super-resolution

Dataset name: multiframe_superres

Data and Inverse Problem

For each frame, the benchmark applies an oriented Gaussian blur, downsamples the blurred image by a factor of two, and adds independent Gaussian noise. With \(K\) frames, the measurements are

\[y_k = S B_k x + n_k, \qquad k=1,\ldots,K,\]

where \(B_k\) is the blur for frame \(k\) and \(S\) is the subsampling operator. Blur orientations are distributed over 180 degrees. The operators are represented as a DeepInv stacked physics object so a solver can process or distribute the frames separately.

The dataset interface fixes the downsampling factor to 2, the blur kernel to 5 x 5, and the blur standard deviation to 1.0. These values are part of the current use case rather than configurable dataset parameters.

What benchopt install Does

multiframe_superres declares no additional dataset-specific software requirements. Installation still resolves requirements from the selected objective and solvers, but it does not download the image. When the repository has already been installed with its Python dependencies, there is normally no extra dataset installation step.

What benchopt prepare Does

Preparation downloads butterfly.png into BenchOpt’s data directory for the multiframe_superres dataset if it is not already cached. Preparation is independent of image size, batch size, noise level, and number of frames, so a single downloaded source image serves every configuration. Preparing on a login node is recommended when compute nodes cannot access the internet.

Available Dataset Parameters

image_size (default 2048)

Target image size. An integer produces a square 2D image.

batch_size (default 1)

Number of copies of the source image in the generated batch.

channels (default 3)

Number of image channels requested from the data interface. This use case is intended for color data.

num_operators (default grid 1, 8, 16)

Number of low-resolution frames and therefore the number of stacked physics operators.

noise_level (default 0.1)

Standard deviation of the Gaussian noise added to every frame.

seed (default 42)

Seed used to initialize the distributed dataset context. Frame noise remains reproducible across runs.

A minimal dataset section is:

dataset:
  - multiframe_superres:
      image_size: 2048
      batch_size: 1
      channels: 3
      num_operators: 8
      noise_level: 0.1
      seed: 42