Characterizing CNN-Based Vessel Detection Algorithm Sensitivity to Optical Sensor Artifacts
Abstract
The number of space-based optical imaging systems have substantially grown in recent years. Many of these commercial based vendors are exploring concepts to move processing applications directly to the space segment or to incorporate advanced algorithms in their ground segment. Typically, Convolutional Neural Network (CNN)-based detection algorithms are considered due to their high reliability in laboratory settings. However, this class of algorithms routinely demonstrates degraded performance when unexpected phenomena are introduced to the data. As more near-real-time applications are considered, CNN-based algorithms will be subjected to sensor calibration artifacts. This paper seeks to characterize the response of CNN-based vessel detection algorithms to common electro-optical sensor calibration artifacts. Several artifacts are explored, including added sensor noise and failed detector artifacts. Each of the explored artifacts uniquely impact the algorithm detection performance. It is found that applying Poisson distributed random noise to the imagery substantially degrades model performance by increasing false alarm rate. Likewise, the application of a uniformly distributed random scale factor to the imagery degrades model performance by lowering positive detection rate. These results may inform plans to implement CNN-based algorithms directly in the space-segment by characterizing typical performance variations due to underlying sensor calibration issues.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 15, 2023
- Accession Number
- AD1210749
Entities
People
- Charles Keene
- John G. Warner
- Michael Tietz
- Quinton Davidson
- William Scharpf
Organizations
- United States Naval Research Laboratory