Operationally Relevant Artificial Training for Machine Learning: Improving the Performance of Automated Target Recognition Systems
Abstract
We set out to demonstrate that an automated target recognition (ATR) system could be built using nothing but commercial off-the-shelf artificial intelligence/machine learning (AI/ML) algorithms and artificially rendered imagery to train thembut we did not succeed. What the AI/ML algorithms learned about detecting artificial objects in artificial images failed to transfer meaningfully toward the task of detecting real objects in real images. However, to our surprise, we discovered that a hybrid training set consisting of both real and artificial images together produced a more robust ATR system than a training set of real images alone. In other words, we were able to boost the performance of the ATR system consistently by adding artificially generated images of the target to the original training set.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 2020
- Accession Number
- AD1115673
Entities
People
- Damien Baveye
- Dara Gold
- Gavin S. Hartnett
- Jasmin Leveille
- Jeff Hagen
- Jia Xu
- Lance Menthe
- Li A. Zhang
Organizations
- RAND Corporation