Efficient and Scalable Visual Computing for Nanoscale Imaging

Author

• Hao Wang, School of ComputingFollow

Date of Award

8-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

School of Computing

Committee Chair/Advisor

Abolfazl Razi

Committee Member

Fatemeh Afghah

Committee Member

Long Cheng

Committee Member

Nianyi Li

Abstract

Nanoscale analysis often relies on instrument-mediated scientific imaging methods to capture visual signals that are difficult to observe, interpret, or quantify through ordinary perception alone. These domains are often characterized by low signal quality, limited annotations, complex morphology, non-natural visual statistics, and strong dependence on physical acquisition processes. This dissertation focuses on efficient visual computing for nanoscale imaging, with an emphasis on the representation, reconstruction, and analysis of high-resolution scientific visual data. In our early work, we relied on rule-based algorithms and conventional machine learning methods for feature extraction. However, these approaches struggle to scale with the increasing complexity and diversity of subvisual data. Despite the rapid development of deep learning, data-driven visual analysis still depends heavily on large-scale annotated datasets, which are rarely available in subvisual imaging. To address data scarcity in subvisual imaging, we integrate generative frameworks, including diffusion-based models, into the proposed representation. These models enable flexible synthesis, reconstruction, and transformation of visual data while preserving diversity and structural consistency. To further support scalability in large-scale settings, we develop both generative pairing and flow-based self-supervised learning strategies. These approaches enable scalable pretraining and support practical deployment in high-throughput imaging scenarios. Building upon these advances, we move beyond traditional pixel-based representations and introduce a fully vectorized visual representation framework based on 2D Gaussian splatting. This approach provides a highly efficient and scalable representation that unifies generation, compression, and reconstruction within a single formulation. Taken together, this dissertation presents a unified framework for structured visual representation, enabling scalable computation in information-dense microscopic and nanoscale imaging domains.

Recommended Citation

Wang, Hao, "Efficient and Scalable Visual Computing for Nanoscale Imaging" (2026). All Dissertations. 4395.
https://open.clemson.edu/all_dissertations/4395

Author ORCID Identifier

0000-0002-3035-3064