Back to search

University of North Texas

Scalable AI: Intelligent Quantization and Neural Architectures for Image Compression, Classification, and Segmentation

Abstract

dc:description

This dissertation explores advanced methodologies in image processing and machine learning, emphasizing the critical role of quantization in optimizing data-driven tasks. It presents a systematic investigation of novel approaches to efficient data representation, compression, and classification, integrating state-of-the-art neural network architectures with traditional signal processing techniques. Chapter 2 focuses on quantization strategies that enhance image compression efficiency while preserving perceptual quality. Innovations such as L1-principal component-based K-means clustering improve cluster initialization, yielding superior SSIM scores. Nonlinear encoding of 3-channel color spaces further optimizes storage and transmission with minimal quality loss. Chapter 3 introduces the OSCAR 2023 compression model, designed for RGB, infrared, and multispectral imagery. It achieves high-fidelity compression, demonstrating strong PSNR and MSSSIM performance across diverse applications, including remote sensing and surveillance. The model’s deployment on Google Coral devices underscores its practicality in edge computing environments. Chapter 4 explores the Efficient Vision Transformer (EViT), a streamlined transformer architecture integrating signal processing techniques to enhance classification tasks. It excels in underwater object detection, COVID-19 diagnosis, and atmospheric disturbance classification on Mars. The CoViT variant achieves peak accuracy in binary and multiclass medical imaging, while EViT’s application in dust storm detection supports future space missions. Chapter 5 extends these advancements with the development of the Multispectral Efficient Land Use Classification (MELC) model and MELC-PC, optimizing EViT for remote sensing applications. Additionally, EViT-like models are adapted for STM32 microcontrollers, achieving real-time inference with minimal computational overhead. The introduction of the EIR-ViT algorithm for infrared target detection further demonstrates the architecture’s adaptability. This work underscores the transformative impact of integrating quantization with modern deep learning models, paving the way for future advancements in artificial intelligence, image compression, and real-time classification. The findings contribute to more efficient, scalable, and high-performance solutions across diverse fields.

Degree

thesis:*
Grantor dc:publisher
University of North Texas
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Depoian, Arthur Charles
Contributors dc:contributor
  • Ashdown, Jonathan
  • Varanasi, Murali R.
  • Hua Sun
  • Dong Xie
  • Namuduri, Kamesh
  • Guturu, Parthasarathy
  • Sos Agaian

Subjects

dc:subject × 6

Rights

dc:rights
Statement dc:rights
  • Public
  • Depoian, Arthur Charles
  • Copyright
  • Copyright is held by the author, unless otherwise noted. All rights Reserved.
Language dc:language
English

Identifiers

dc:identifier.*
Identifier
local-cont-no: submission_4266
https://digital.library.unt.edu/ark:/67531/metadc2443141/
ark: ark:/67531/metadc2443141

Chain of custody

source
Harvested from
University of North Texas
Base URL
digital.library.unt.edu/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Depoian, Arthur Charles. Scalable AI: Intelligent Quantization and Neural Architectures for Image Compression, Classification, and Segmentation. University of North Texas, 2025. https://doi.org/10.12794/metadc2443141