How mingc0816/f2dm works
F2DM is an underwater-image enhancement system built on BasicSR. Its documented workflow takes paired training images or underwater inputs, runs frequency-domain restoration, and produces enhanced images; training also evaluates restoration quality. The README describes wavelet decomposition, separate frequency-band modulation, and cross-frequency fusion. The tree and excerpts support the framework’s training, model, data, loss, and metric layers, but do not include the README’s referenced test entry point or configuration files, so testing orchestration is not mapped.
Main entry points: Wfdiff_model.py F2DM model · train.py Training runner · single_image_dataset.py Single-image data · wfdiffx2_arch.py Enhancement network
Components 16
Training workflow
- Training runner
basicsr/train.py - F2DM model
basicsr/models/Wfdiff_model.py
Image data
- Paired datasets
basicsr/data/paired_image_dataset.py - Single-image data
basicsr/data/single_image_dataset.py - Image transforms
basicsr/data/transforms.py - File access
basicsr/utils/file_client.py
F2DM restoration
- Enhancement network
basicsr/archs/wfdiffx2_arch.py - Wavelet decomposition
basicsr/models/wavelet.py - Band modulation
basicsr/archs/Padiff_arch/dft_arch.py - Cross-band fusion
basicsr/archs/Padiff_arch/dft_arch.py
BasicSR services
- Model base
basicsr/models/base_model.py - Training losses
basicsr/losses/losses.py - Image metrics
basicsr/metrics/psnr_ssim.py
Other components
- Researcher
- Enhanced images
- Reference images
Connections 19
- Researcher → Training runner: starts training
- Researcher → Single-image data: provides inputs
- Paired datasets → File access: reads image pairs
- Single-image data → File access: reads images
- Paired datasets → Image transforms: transforms pairs
- Single-image data → Image transforms: transforms images
- Training runner → Paired datasets: builds datasets
- Training runner → F2DM model: optimizes model
- Training runner → Image metrics: evaluates outputs
- F2DM model → Model base: uses model utilities
- F2DM model → Enhancement network: runs enhancement
- Enhancement network → Wavelet decomposition: decomposes bands
- Enhancement network → Band modulation: modulates frequencies
- Enhancement network → Cross-band fusion: fuses frequencies
- F2DM model → Training losses: computes training loss
- Paired datasets → Reference images: supplies references
- Image metrics → Reference images: compares references
- F2DM model → Enhanced images: produces enhanced images
- Single-image data → F2DM model: provides input images