nekhtiari/image-similarity-measures

Image Similarity Measures

What it is – A small Python library (and accompanying CLI) that computes how alike two images are. It implements eight classic similarity/quality metrics such as RMSE, PSNR, SSIM, FSIM, and a few others that are common in remote‑sensing and computer‑vision research.

Why it matters – When you train or test an image‑processing model (e.g., super‑resolution, denoising, or satellite‑image reconstruction) you need a quantitative way to say whether the output looks like the ground‑truth. This package gives you those numbers in a ready‑to‑use form, both from Python code and from the command line.

How to get it – Install with pip (Python ≥ 3.10):

pip install image-similarity-measures

Optional speed‑up for the FSIM metric (pyfftw) or TIFF handling (rasterio) can be added via extras:

pip install image-similarity-measures[speedups]
pip install image-similarity-measures[rasterio]

Command‑line usage – Compare two files and output a JSON object with the selected metrics:

image-similarity-measures \
  --org_img_path=a.tif \
  --pred_img_path=b.tif \
  --metric=rmse --metric=psnr   # repeat --metric for more, or use "all"

The tool expects channel‑last arrays (height × width × channels).

Python usage – Call the high‑level evaluation helper or any metric directly:

from image_similarity_measures.evaluate import evaluation
results = evaluation(
    org_img_path="example/lafayette_org.tif",
    pred_img_path="example/lafayette_pred.tif",
    metrics=["rmse", "psnr"]
)

# Or a single metric
from image_similarity_measures.quality_metrics import rmse
score = rmse(org_img, pred_img)

Extensibility – The repository invites contributions (see README-dev.md for developer guidelines).

Citation – If you use the code in a publication, cite the authors’ 2020 ISPRS Annals paper on super‑resolution of multispectral satellite images.

---

All details above are taken directly from the project's README.