DenoisingMetrics

This is a codebase for evaluating Image Denoising and Enhancement tasks, supporting both reference-based and no-reference metrics.

Metrics

• Reference-based: PSNR, SSIM, LPIPS, FID
• No-reference: UIQM, UCIQE (Underwater Image Quality/Colorfulness Measures)

Requirements

Install dependencies using:

pip install -r requirements.txt

Key dependencies: torch, numpy, scipy, opencv-python, scikit-image, tqdm, lpips, pytorch_fid, threadpoolctl.

---

1. Batch Evaluation (main.py)

Use main.py to evaluate multiple methods across multiple datasets at once using a configuration file.

Usage

python main.py --config config.json

Configuration (config.json)

The configuration file allows you to define global settings and dataset-specific paths:

• device: cuda or cpu.
• all_metrics: If true, computes UIQM and UCIQE even if Ground Truth is available.
• output_path: Directory where CSV results and logs will be saved.
• datasets: A dictionary where each key is a dataset name, containing:
  • parentFolder: Root path for the dataset.
  • gt_dir: Subfolder containing ground truth images.
  • fid_gt_cache: (Optional) Path to a .npz file containing precomputed FID statistics for the GT.
  • pred_dir: A dictionary mapping Method Names to their respective result subfolders.

---

2. Direct Folder Comparison (compute.py)

Use compute.py for a quick comparison between two specific folders (Predicted vs. Ground Truth).

Usage

python compute.py -p /path/to/predicted_images -g /path/to/ground_truth_images

Features:

• Auto-device detection: Automatically uses CUDA if available.
• Resource protection: Limits BLAS threads during FID calculation to prevent system hangs.
• Speed: GPU-accelerated LPIPS and FID feature extraction.

---

3. No-Reference Evaluation (unref.py)

Use unref.py to calculate UIQM and UCIQE for a single folder of images without requiring ground truth.

Usage

python unref.py --folder /path/to/images --output results.txt

---

Parameters Summary

Script	Parameter	Description
main.py	--config, -c	Path to the JSON configuration file.
compute.py	--pred_dir, -p	Directory of predicted images.
compute.py	--gt_dir, -g	Directory of ground truth images.
compute.py	--device	Device to use (auto/cpu/cuda).
compute.py	--strict_check, -s	Ensure filenames match exactly between folders.
unref.py	--folder, -f	Directory of images to evaluate.
unref.py	--output, -o	File path to save the per-image and average results.