Find the blurry, dark, and duplicate images hiding in your dataset

What it does

CleanVision scans collections of image files and flags common quality defects: exact or near duplicates, blurry or low-entropy shots, under- and over-exposed frames, grayscale images, and odd aspect ratios or sizes. It generates a report of these issues, meant as a quick sanity check before you start training computer vision models. The tool works on standard image folders and is positioned as a companion to Cleanlab’s label-auditing package.

The interesting bit

The project treats dataset hygiene as a first-class step rather than an afterthought, inspecting raw pixels for visual defects instead of annotation errors. It is deliberately narrow: it will not fix your labels or catch a misclassified cat, but it will spot if half your “cats” are pitch-black thumbnails.

Key highlights

• Detects nine predefined issue types, from exact duplicates to odd aspect ratios
• Runs on standard image folders as well as HuggingFace and Torchvision datasets
• Supports Linux, macOS, and Windows on Python 3.10+
• Produces a structured report that can be filtered to specific issue types
• Focuses purely on raw image quality; label issues are explicitly out of scope (handled by the separate cleanlab package)

Verdict

Worth a look if you are sitting on a messy directory of scraped or collected images and need automated triage before labeling or training. Skip it if your data is already pristine or if your problem is label noise rather than image quality.