pyteenybrisque

Tiny BRISQUE no-reference image quality scorer. One function, two runtime dependencies (numpy and Pillow), ~250 KB of vendored model weights.

import pyteenybrisque

score = pyteenybrisque.score(image="photo.jpg")
print(score)  # lower is better; ~0-100 scale

score() accepts a path, a PIL.Image.Image, or a numpy array (HxW grayscale or HxWx{3,4} RGB / RGBA, uint8 or float in [0, 1]).

Installation

pip install pyteenybrisque

What it computes

BRISQUE (Mittal, Moorthy, Bovik 2012) is a no-reference image quality metric. It extracts 36 natural-scene-statistics features from the luma channel at two scales and runs them through an RBF SVR trained on LIVE IQA. Lower scores mean higher perceived quality.

The implementation matches pyiqa's BRISQUE within ~0.1 BRISQUE points on natural images.

How it compares

Each metric in the table below was scored on the Kodak True Color test set (8 lossless 768×512 PNGs) under six degradation sweeps. Per source and metric, scores are min-max normalised across the sweep so 0 = best in run, 1 = worst; the line is the median across sources, the shaded band is the inter-quartile range.

BRISQUE is competitive with the deep-learning metrics on every degradation. The benchmark script lives at tools/benchmark_metrics.py and is reproducible end-to-end.

Why "teeny"

pyiqa is the right tool if you want every IQA metric in one place. It pulls in PyTorch and ~2 GB of dependencies. This package does one metric, on top of just numpy and Pillow, in ~250 KB. Use it when BRISQUE is all you need.

License

MIT