Analysis Methodology of Image Classifiers in Stressed Environments
Abstract
This report defines the methodology used to analyze 2D (image) classifiers in the form of convolutional neural network models in the presence of battlefield stresses. Battlefield stresses are defined in this context as those that alter the sensor data relative to the images used to train the neural networks. Battlefield stresses can be intentional, such as camouflage or distortion, or unintentional, such as obscurants or excessive background (or foreground) clutter. Parts needed to perform the analysis are described first, followed by the methodology steps used to rank and rate the neural network classifiers, and the report concludes with a practical analysis case.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 2022
- Accession Number
- AD1180069
Entities
People
- Patrick S. Debroux