CleanVision: Automated Image Dataset Auditing for Computer Vision
CleanVision is a Python library developed by Cleanlab that automates the detection of common issues in image datasets before they are used for training machine learning models. Released in 2022 under the Apache 2.0 license, it serves as a critical tool for ensuring data and training integrity in computer vision projects. The library is designed to identify problems such as duplicates, blurry images, exposure issues, and other anomalies that could compromise model performance.
The tool operates as a data-centric solution, providing a quick and efficient way to audit datasets. CleanVision is particularly useful for tasks including classification, segmentation, object detection, pose estimation, keypoint detection, and generative modeling, where high-quality data is essential.
CleanVision’s functionality is built around the Imagelab class, which allows users to specify a dataset path and automatically scan for predefined issues. The library generates detailed reports that highlight problematic images, enabling users to address these issues before proceeding with model training. This approach reduces the risk of training models on suboptimal data, which can lead to poor performance or biased outcomes.
How It Works
CleanVision simplifies the process of dataset auditing through a streamlined workflow. Users begin by installing the library via pip and then initialize an Imagelab instance by specifying the path to their dataset. The library then scans the images for a predefined set of issues, including exact duplicates, near duplicates, blurry images, low-information content, exposure problems, grayscale images, and unusual aspect ratios or sizes. Each issue type is associated with a specific detection mechanism, ensuring comprehensive coverage of potential data problems.
For example, exact duplicates are identified by comparing pixel-level data, while near duplicates use visual similarity metrics. This adaptive approach ensures that CleanVision remains effective across diverse datasets.
Once the audit is complete, CleanVision generates a report that categorizes issues by type and provides examples of affected images. Users can also customize the audit by specifying only certain issue types to focus on, such as dark or blurry images. The library’s output is designed to be actionable, with clear indicators of which images require attention.
Use Cases
CleanVision is applicable across a broad spectrum of computer vision tasks where data quality directly impacts model performance. In image classification projects, for instance, the presence of duplicates or low-information images can skew training, leading to overfitting or poor generalization. By identifying and removing such issues, CleanVision helps ensure that models are trained on a representative and high-quality dataset.
For segmentation and object detection tasks, CleanVision’s ability to detect blurry or poorly exposed images is invaluable. Similarly, in generative modeling, where datasets often include synthetic or augmented images, CleanVision helps maintain consistency by identifying anomalies that could distort the training process.
Many large-scale datasets are compiled from multiple sources, increasing the likelihood of inconsistencies such as varying aspect ratios or duplicate entries. CleanVision’s automated auditing reduces the manual effort required to identify these issues, making it a scalable solution for teams working with extensive image collections.
Key Features & Capabilities
CleanVision’s core functionality revolves around detecting nine specific types of image issues, each addressed through specialized algorithms:
- Exact Duplicates: Identifies images that are identical at the pixel level, which can inflate dataset size without adding useful information.
- Near Duplicates: Flags images that are visually similar but not identical, often resulting from minor variations in capture conditions or post-processing.
- Blurry Images: Detects images where details are indistinct due to motion or focus issues.
- Low Information: Highlights images with minimal content or high entropy in pixel values, which may not contribute meaningfully to training.
- Dark and Light Images: Flags under-exposed or over-exposed images, which can distort feature recognition.
- Odd Aspect Ratio: Identifies images with unusual width-to-height ratios, which can cause issues in resizing or cropping during preprocessing.
- Odd Size: Flags images that are abnormally large or small compared to the rest of the dataset, which may indicate corruption or mislabeling.
The library’s design emphasizes usability, with minimal setup required to begin auditing a dataset. Users can install CleanVision via pip and run a few lines of Python code to initiate an audit. The Imagelab class provides methods to specify issue types, generate reports, and visualize findings, making it accessible to both beginners and experienced practitioners.
Version 0.3.7 updated the license to Apache 2.0, making it more permissive for broader adoption. This release also refined output handling, allowing users to control verbosity and focus on specific issues.