Skip to content

Image Corruption Toolbox

(aiverify.stock.image-corruption-toolbox) [source]

Description

This plugin tests the robustness of AI models to natural corruptions.

There are four different broad groups of corruptions that are packaged in this plugin. Each of these broad groups of corruptions also have more specific corruption functions indicated in brackets below:

  • General (Gaussian, Poisson, Salt and Pepper)
  • Blur (Gaussian, Glass, Defocus, Horizontal Motion, Vertical Motion, Zoom)
  • Digital (Brightness Up and Down, Contrast Up and Down, Saturate Up and Down, Random Perspective, JPEG Compression)
  • Environmental (Snow, Fog, Rain)

The toolbox generates corrupted images based on the uploaded test data at 5 different severity levels for each corruption function. The accuracy of the model is calculated with the new corrupted datasets.

Plugin Content

  • Algorithms
Name Description
Blur Corruptions Algorithm that adds blur corruptions (gaussian, glass, defocus, horizontal motion, vertical motion, zoom) to images across thresholds of interests, and calculates the accuracy of the model
Digital Corruptions Algorithm that adds digital corruptions (brightness up and down, contrast up and down, saturate up and down, random perspective, jpeg compression) to images across thresholds of interests, and calculates the accuracy of the model
Environment Corruptions Algorithm that adds environmental corruptions (snow, fog, rain) to images across thresholds of interests, and calculates the accuracy of the model
General Corruptions Algorithm that adds general corruptions (gaussian, poisson, salt and pepper) to images across thresholds of interests, and calculates the accuracy of the model
  • Widgets
Name Description
Introduction To provide an introduction to the Image Corruption Toolbox
Understanding Bar Chart To guide your users on reading the generated bar charts
Bar Chart (by corruptions type) To generate bar chart to visualise the accuracy results when blur corruptions are applied
  • Blur corruption samples
  • Digital corruption samples
  • Environment corruption samples
  • General corruption samples
 To generate sample images for the different corruption types
Recommendation To provide recommendations for robustness (image corruptions) testing

Algorithm Details

Refer to the README file under each algorithm for more details on the algorithm.