Author ORCID Identifier

https://orcid.org/0000-0002-0110-0949

Semester

Fall

Date of Graduation

2025

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Jeremy Dawson

Committee Member

Matthew Valenti

Committee Member

Brian Powell

Committee Member

Prashnna Gyawali

Committee Member

Jacqueline Speir

Abstract

Despite the recent expansion of machine learning algorithms to cover a wide range of disciplines, several areas of automatic target recognition (ATR) remain underexplored. This dissertation presents tools developed to improve performance in three significant aspects of ATR: semi-supervised annotation, sensor fusion, and image super-resolution. The aim of the semi-supervised methods is to automatically annotate targets in scenarios where labeled data are scarce in the target domain but available in the source domain. Secondly, to address the limitations of individual image sensors and enhance robustness under different environmental conditions and man-made constraints, a sensor fusion algorithm was developed to improve classification performance. Furthermore, since aerial, spaceborne, and ground-based target images are often degraded by noise, motion blur, and atmospheric turbulence, an effective ATR system is necessary to improve image quality toward the ground-truth manifold. To this end, a multi-granular perceptual objective is introduced to facilitate the super-resolution of degraded images, producing more realistic and higher-quality image reconstructions.

Within the semi-supervised annotation framework, a novel domain translation method is applied that effectively transfers knowledge from the source to the target domain while preserving class information via a pretrained classifier, addressing the inherent one-to-many ambiguity of unpaired domain translation functions. The annotation approach produces more consistent and accurate results than several well-known semi-supervised techniques for target recognition. Furthermore, a multi-sensor target classification method is introduced that can capture diverse representations—including low- and high-frequency features, fine structural details, and heterogeneous characteristics across different image sensors—thereby reducing granularity gaps and enhancing robustness under various constraints. This approach outperforms several state-of-the-art algorithms for target classification. Finally, a novel family of perceptual loss functions for image super-resolution that preserves information through a lossless invertible neural network trained on ImageNet is presented. This objective significantly improves both diffusion- and GAN-based super-resolution methods, enhancing their robustness to real-world degradations. Comprehensive visualizations and quantitative evaluations demonstrate that the proposed methods yield superior performance across multiple aspects of automatic target recognition.

Download

Share

COinS