RetinaHash

Lightweight, pure-NumPy perceptual hashing for images

RetinaHash computes compact, rotation-invariant fingerprints of images using only NumPy...no OpenCV, no PIL, no Scipy.
Perfect for duplicate detection, near-duplicate search, and lightweight image clustering.

Features

• Pure NumPy – runs anywhere Python runs
• Rotation invariant – same hash for 0°, 90°, 180°, 270° rotations
• Auto-crop – strips solid scanner borders automatically
• Extensible – plug-in architecture for new hash algorithms
• Fast – ~1 ms per 512×512 image on a laptop CPU (TRUST ME BRO Benchmark, lol.)

Install

pip install retinahash

Or better, because you are sane and use uv to manage your projects:

uv add retinahash

(or drop the retinahash/ folder into your project—zero dependencies beyond NumPy).

Quick Start

import imageio.v3 as iio          # any loader will do but numpy array is expected
from retinahash import RetinaHash

img = iio.imread("photo.jpg")
h   = RetinaHash.phash(img, rotation_invariant=True)
print(h)        # 16-char hex string, e.g. 'a5f3c1e8d2b4906e'

Compare two images (via Hamming distance):

h1 = RetinaHash.phash(img1)
h2 = RetinaHash.phash(img2)
distance = RetinaHash.distance(h1, h2)

API

RetinaHash.hash(image, method='phash', hash_size=8, rotation_invariant=False, auto_crop=False)
RetinaHash.phash(image, **kwargs) – convenience wrapper
RetinaHash.distance(hex1, hex2) – Hamming distance between two hex hashes

Algorithms

More algorithms (ahash, dhash, whash, …) can be added by dropping a new module into retinahash/strategies/ and registering it in STRATEGIES.

Customisation

# 16-bit hash, auto-crop on, rotation invariant
h = RetinaHash.hash(img, method='phash', hash_size=16,
                    auto_crop=True, rotation_invariant=True)

License

MIT – do what you want, just keep the copyright notice. See LICENSE file.

That’s it—clone, hash, dedupe.