Summary:
UCLA's Department of Mathematics researchers have developed an innovative method for the rapid and precise interpretation of SAR imagery, setting a new standard in image analysis efficiency.
Background:
Synthetic aperture radar (SAR) is utilized as an image reconstruction method for a wide array of applications, including satellite monitoring, to capture fine resolution images. Increasingly, there is a need for automatic target recognition (ATR) of the objects within SAR images so that users can avoid the need to manually label objects in images. One potential solution is the use of supervised machine learning, relying on abundant labeled data, which can be costly and unfeasible. Active machine learning allows for the selection of specific unlabeled data points that are subsequently labeled by a human expert, improving classification with fewer data points. Labeling and classifying SAR data rely on either graph-based or neural network-based methods. Graph-based methods are complicated by extraneous information that is present in large datasets consisting of raw images of scenes. Neural network methods allow for effective extraction of patterns from SAR images but require significant training data to avoid overfitting, which is often unavailable. There remains an unmet need for an active learning method for the labeling and classifying of SAR data and automatic target recognition.
Innovation:
UCLA researchers in the Department of Mathematics developed a novel method for improving the classification SAR data. This technology integrates elements of graph-based learning and neural network methods within an active learning framework. Features are extracted from SAR data by mathematical transformations that reduce dimensionality and capture meaningful patterns. Similarity plots are subsequently constructed from the embedded data, representing relationships between points. Both labeled and unlabeled data are then classified using a graph-based semi-supervised learning approach, reducing the risk of overfitting. This technology has successfully demonstrated classification of SAR data when a small amount of labeled data is accessible and improves generalization performance.
Potential Applications:
- Remote sensing
- Infrastructure monitoring
- Surveillance
- Urban planning
- Oceanography & environmental monitoring
- Geological monitoring
- Military surveillance and astronomy
Advantages:
- Reduced labeling costs
- Optimized data selection
- Flexibility
- Performance improvement
- Cost-effective
Development To Date:
First successful demonstration of technology is complete.
Reference:
UCLA Case No. 2022-175
Lead Inventor:
Andrea Bertozzi
Relevant Publications:
Graph-based Active Learning for Semi-supervised Classification of SAR Data, draft of manuscript accepted to 2022 SPIE Defense and Security Conference, Jeffrey Calder, Kevin Miller, John Mauro, Jason Setiadi, and Andrea L. Bertozzi.
Y. Qiao, C. Shi, C. Wang, H. Li, M. Haberland, X. Luo, A. M. Stuart, and A. L. Bertozzi, Uncertainty Quantification for Semi-supervised Multi-class Classification in Image Processing and Ego-Motion Analysis of Body-Worn Videos, IS&T International Symposium on Electronic Imaging 2019: Image Processing: Algorithms and Systems XVII proceedings, Burlingame, CA.Hao Li, Honglin Chen, Matt Haberland, Andrea L. Bertozzi, and P. Jeffrey Brantingham.
PDEs on graphs for semi-supervised learning applied to first-person activity recognition in body worn video, DCDS, online first, 2021.Zhaoyi Meng, Javier Sanchez, Jean-Michel Morel, Andrea L Bertozzi, P. Jeffrey Brantingham.
Ego-motion Classification for Body-worn Videos, 2018 Proceedings of the 2016 International Conference on Imaging, vision and learning based on optimization and PDEs, Bergen, Norway, pp. 221239.
Kevin Miller and Andrea L. Bertozzi, Model-Change Active Learning in Graph-Based Semi-Supervised Learning, submitted to SIAM J. on Mathematics of Data Science, 2021.