National Institute of Standards and Technology

Feature Sensitivity Study

Summary

Our work addressed the following questions: (1) Do feature parameter significantly impact the extracted feature values?; (2) What are the most important algorithmic factors of image feature extraction affecting the derived feature values?

A total of 9 independent parameters have been identified. They can be categorized in:

  • 6 procedural parameters (i.e., the ones involving an algorithmic choice)
  • 3 robustness parameters
  • An extra 2 parameters, that directly depend on the previous 9 parameters (i.e. they do not add any degrees of freedom)

A total of 75 features have been computed over 1152 images of the Big Data experiment, producing over a billion feature values. The output is stored in a 12GB sqlite database, composed of 18.6 million rows.

Here below is a visualisation of the impact of 2 of the 9 parameters: the tiling shape and size. The shape of the tile can be Hexagonal or Square, and the size of the tiles can Small, Medium or Large. Shown in Figure 1, is an illustrated subset with a total of 2 shapes x 3 sizes = 6 combinations:

Visualisation of the impact of the tiling shape and size on the extracted features
Visualisation of the impact of the tiling shape and size on the extracted features

Description

The details of the 9(+2) independent and dependent parameters explored in the sensitivity analysis are listed here below:

Detail of the procedural, robustness and dependent parameters being studied
Detail of the procedural, robustness and dependent parameters being studied

Date created: April 10, 2014 | Last updated: