vision-data-audit 0.1.2

Quick audit of an image folder: formats, sizes, mean/std, duplicates.

vision-data-audit

Audit an image dataset folder in seconds. Get counts, stats, and duplicates.

What you get

• Image formats (JPEG, PNG, etc.)
• Color modes (RGB, grayscale, etc.)
• Size distribution (width/height percentiles)
• RGB mean/std (quick estimate for normalization)
• Duplicate detection (perceptual hash)

Install

pip install vision-data-audit

Usage

Basic audit:

vision-data-audit /path/to/images

Save to JSON:

vision-data-audit /path/to/images --out report.json

Skip duplicates (faster):

vision-data-audit /path/to/images --no-dupes

Options

• --out FILE - save report as JSON
• --no-dupes - skip duplicate detection
• --sample N - max images for mean/std (default: 200)

Python

from vision_data_audit.core import audit_folder

report = audit_folder(
    "/path/to/images",
    sample_images=200,
    hash_size=8,
    compute_dupes=True,
)

print(report["formats"])
print(report["width"], report["height"])
print(report["mean_std_rgb_approx"])

# Duplicate groups (if enabled)
print(report.get("duplicates_dhash", []))

Notes

• Mean/std uses a random sample for speed
• Duplicates found using perceptual hashing (dHash)

Help improve this

If you use this tool, please open a GitHub issue with:

• the command you ran
• the printed summary (or report.json with paths removed)
• what you wish it reported (corrupt files? split leakage? near-duplicates?)

Ideas on the roadmap:

• near-duplicate grouping by Hamming distance threshold
• train/val/test leakage detection across subfolders
• faster mean/std computation (vectorized)

Development

uv sync
uv run vision-data-audit --help
uv run pytest
uv build

License

Apache-2.0