Electrical and Computer Engineering, Department of
First Advisor
Benjamin S. Riggan
Committee Members
Michael Hoffman, Mohammad Hasan
Date of this Version
7-2025
Document Type
Thesis
Citation
A thesis presented to the faculty of the Graduate College at the University of Nebraska in partial fulfillment of requirements for the degree of Master of Science
Major: Electrical Engineering
Under the supervision of Professor Benjamin S. Riggan
Lincoln, Nebraska, July 2025
Abstract
Atmospheric turbulence presents a significant barrier to long-range facial recognition, introducing severe geometric distortions and blur that degrade image quality. This thesis investigates deep learning approaches for mitigating these effects, with a focus on transformer based architectures and domain adaptation strategies.
An in-depth benchmarking study was performed using convolutional neural networks (CNNs) and vision transformers (ViTs) on the Husker BRIAR Research Collection from up to 500m (HBRC-500) face dataset. The results demonstrated that vision transformers, particularly hierarchical vision transformers like the shifted-window (Swin) transformer, outperform CNN-based models at long distances due to their ability to model global spatial relationships and compensate for turbulence-induced distortions.
Further experiments were conducted with the HBRC-500 and Biometric Recognition and Identification at Altitude and Range (BRIAR) datasets and a key contribution of this work is the development of deformable patch embeddings (DPE), which incorporate a lightweight offset network to adaptively correct geometric distortions as part of the tokenization process. DPE, trained with synthetic turbulence generated from both random and Kolmogorov-based simulations, yielded substantial improvements in Rank-1 accuracy and verification metrics, particularly at extreme ranges. This method proved more effective than generative restoration techniques like denoising diffusion probabilistic models (DDPM), as presented in AT-DDPM, which struggled with identity preservation and domain mismatch, and outperformed weakly supervised domain adaptation strategies that degraded gallery image quality. The DPE solution achieved competitive results with the image-quality informed low-rank adaptation (LoRA) framework of PETALface when evaluated with the same WebFaces4M pretrained backbone.
Additionally, this thesis introduced a novel attention rollout framework for hierarchical Swin transformers, offering high-resolution visualizations of model attention patterns that surpassed traditional Grad-CAM in interpretability. Attention rollout had previously only been implemented for columnar vision transformer architectures and did not account for the effect that windowed attention and patch merging had on the raster scan ordering of tokens when combining attention scores across layers.
Collectively, the findings underscore the viability of transformer-based backbones combined with spatially adaptive sampling mechanisms for long-range face recognition. The proposed methods close a significant portion of the domain gap between pristine indoor enrollment and degraded outdoor probes without relying on generative image restoration. This research lays the groundwork for future efforts in scalable, domain-robust biometric recognition systems under adverse imaging conditions.
Advisor: Benjamin S. Riggan
Included in
Artificial Intelligence and Robotics Commons, Computer Engineering Commons, Electrical and Computer Engineering Commons
Comments
Copyright 2025, Zachary Michael Swanson. Used by permission