Electronic Theses and Dissertations

Date

2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Electrical & Computer Engineering

Committee Chair

Aaron Robinson

Abstract

This dissertation synthesizes a method based on studies focused on enhancing segmentation and inpainting of obscured regions within thermal imagery, specifically targeting cloud segmentation in Long-Wave Infrared (LWIR) images. The research extensively evaluates various classical segmentation algorithms, including thresholding, edge-based, region-based, clustering, and texture analysis. We extensively studied the performance of these methods. The findings from this extensive study helped us decide where to take this research. The next research paper introduces a new approach to the tile-based segmentation of occlusion clouds using the Gray Level Co-Occurrence Matrix (GLCM). This technique extracts features from local tiles to train a classifier, achieving inter-region homogeneity and accuracy performance compared to traditional methods such as Gabor segmentation or Markov Random Field (MRF) schemes. The GLCM-based method distinguishes between dust cloud tiles and clear tiles, facilitating conditional processing crucial for autonomous vehicle navigation and security surveillance applications, even in varying lighting conditions. Additionally, this work explores the use of Generative Adversarial Networks (GANs) for occlusion inpainting in environments compromised by smoke, fog, and other atmospheric obscurants—a common challenge in surveillance, security, and defense. The process involves obscured region segmentation followed by GAN-based pixel replacement, leveraging architectures similar to Pix2Pix and UNet for enhanced image-to-image translation and segmentation tasks. This approach restores obscured sections efficiently, significantly improving the precision and speed of the inpainting process. Integrating these advanced segmentation and inpainting techniques through refined deep-learning architectures marks a significant advancement in thermal imaging technology. It mitigates critical information loss due to atmospheric interference and enhances the analytic capabilities of thermal imaging systems. Future work will expand these models to provide real-time, multi-label semantic segmentation, potentially transforming critical real-world applications. This research underscores a significant leap forward in the capabilities of thermal imaging technologies, setting the stage for broadened applications and improved utility in challenging visual conditions.

Comments

Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to ProQuest.

Notes

Open access.

Share

COinS