Feature Extraction from Large-Scale Complex and Imperfect Data

Abstract

Modem surveillance, target acquisition, and reconnaissance systems are generating terabytes of images and videos continuously. Extracting useful information from the large amounts of data is a challenging but important task to improve a commander s situational awareness and the consequent decision making. Moreover, in military applications, the measurements are often obtained in an extreme environment and suffer from quality degradation significantly. Motivated by object recognition and event identification from surveillance data, this proposal aims to extract features from high-dimensional, complex, and imperfect data by exploiting low-dimensional structures. One research objective is to develop a unified dictionary learning approach when the measurements contain data losses, erroneous measurements, and are low-resolution. The other research objective is to develop new learning methods based on innovative low-dimensional models of the data. This proposal will develop a framework of learning methods when the signals have complex structures, and the obtained data suffer from quality issues. One distinctive feature of this proposal is the quantitative analysis of the learning performance of the proposed methods, especially the dependence on various parameters characterizing data imperfectness and data complexity. The developed methods and tools, together with their corresponding theoretical analysis, contribute to the development of the theory and the practice of high-dimensional data analysis using low-dimensional models. The proposed methods have broad applications in feature extraction, object detection and recognition, image classification, and image denoising. This proposal will develop new models to characterize the intrinsic structures in the practical datasets. The models extend the current model and describe both the strong and the weak c01relations in the data. By connecting low-dimensional structures with the Hankel matrix, the new model further exploits the temporal correlations, which are often ignored in the literature. These fine models serve as a solid basis for the development of new tools that are tailored for these structures to enhance the learning performance. This proposal will develop computationally efficient dictionary learning, classification, and subspace clustering methods with analytical performance guarantees. The low-dimensional structures are captured by non-convex constraints, which are often relaxed into convex constraints in the literature to enable efficient algorithms with performance guarantees. Non-convex-optimization-based approaches have better numerical performance than their convex alternatives, while the theoretical analysis of non-convex approaches is notably lagging. Besides analyzing convex methods, this proposal will mainly focus on the design and the analysis of non-convex approaches. The project abstract is publicly releasable.

Document Details

Document Type
DoD Grant Award
Publication Date
Sep 11, 2018
Source ID
W911NF1710407

Entities

People

  • Meng Wang

Organizations

  • Army Contracting Command
  • Rensselaer Polytechnic Institute
  • United States Army

Tags

Fields of Study

  • Computer science
  • Engineering

Readers

  • Computational Modeling and Simulation
  • Distributed Systems and Data Platform Development
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Machine Learning Algorithms