Scale Equivariance Regularization and Feature Lifting in High Dynamic Range Modulo Imaging

Open In Colab

This repository implements ModuloSEFLnet, a deep learning model for High Dynamic Range (HDR) imaging reconstruction from modulo measurements. The approach combines:

• Feature Lifting: Extracts gradient features from modulo measurements to enhance reconstruction
• Scale Equivariance Regularization: Enforces scale-invariant properties during training for robust HDR recovery
• DRUNet Architecture: Leverages a U-Net backbone with residual blocks for image reconstruction

The model tackles the inverse problem of recovering HDR images from modulo wrapping artifacts caused by sensor saturation at a threshold value.

• Physics Model: Spatial unwrapping with configurable threshold (default: 1.0) and floor mode
• Network: Modified DRUNet with feature lifting (gradient-based) input processing
• Training Loss: MSE loss + scale equivariance regularization
• Metrics: For the demo we report PSNR and SSIM as baseline metrics; however, we recommend using HDR-specific perceptual metrics for a more accurate assessment of HDR image quality:
  • HDR-VDP [1, 2]: perceptual difference metric designed for HDR content (compute on linear HDR radiance or converted to scene-referred units).
  • PU21 [3]: convert HDR to the PU21 perceptually-uniform domain and report PU-PSNR / PU-SSIM for more reliable fidelity measures.

This project requires the deepinv library for physics modeling, network architectures, and metrics.

Install dependencies:

pip install deepinv torch torchvision matplotlib numpy tqdm

Train the ModuloSEFLnet model on the UnModNet HDR dataset:

python train.py

The training script automatically evaluates on the test set after each epoch and reports PSNR/SSIM metrics.

Test the trained model and visualize results:

python test.py

This loads the trained checkpoint from ./ckpts/ModuloSEFLnet.pth.

Expected dataset format:

data/
 unmodnet/
 source/
 *.npy files (training images)
 unmodnet_test/
 source/
 *.npy files (test images)

Images should be stored as .npy files and will be normalized to the range [0, MAX_VALUE].