Unresolved Object Detection Using Synthetic Data Generation and Artificial Neural Networks
Abstract
This research presents and solves constrained real-world problems of using synthetic data to train artificial neural networks (ANNs) to detect unresolved moving objects in wide field of view (WFOV) electro-optical/infrared (EO/IR) satellite motion imagery. Objectives include demonstrating the use of the Air Force Institute of Technology (AFIT) Sensor and Scene Emulation Tool (ASSET) as an effective tool for generating EO/IR motion imagery representative of real WFOV sensors and describing the ANN architectures, training, and testing results obtained. Deep learning using a 3-D convolutional neural network (3D ConvNet), long short term memory (LSTM) network, and U-Net are used to solve the problem of EO/IR unresolved object detection. U-Net is shown to be a promising ANN architecture for performing EO/IR unresolved object detection. In two of the experiments, U-Net achieved 90 percent and 88 percent pixel prediction accuracy. In addition, the results show ASSET is capable of generating sufficient information needed to train deep learning models.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 2019
- Accession Number
- AD1076436
Entities
People
- Yong U. Sinn
Organizations
- Air Force Institute of Technology