Development of deep generative imaging technique of capacitive resistivity data for underground object detection

Date of Publication

2024

Document Type

Master's Thesis

Degree Name

Master of Science in Electronics and Communications Engineering

Subject Categories

Electrical and Computer Engineering | Engineering

College

Gokongwei College of Engineering

Department/Unit

Electronics And Communications Engg

Thesis Advisor

Argel A. Bandala

Defense Panel Chair

Jose Martin Z. Maningo

Defense Panel Member

Ryan Rhay P. Vicerra
Ronnie S. Concepcion II

Abstract/Summary

Non-invasive underground imaging through the capacitive resistivity technique allows measuring the electrical resistivity of the subsurface and requires various imaging processing techniques to map the subsurface. However, the reconstruction of the subsurface resistivity images needs an iterative process and lengthy, complex mathematical approaches, resulting in ill-posedness and poor-quality images. Although different advancements in deep generative AI for image synthesis continue to be a dynamic and rapidly evolving field, none of these were explored for capturing nuances from underground imaging data to generate resistivity images.

With that, this study aims to develop a deep generative imaging technique based on a Generative Adversarial Network (GAN) for underground object detection by learning to map apparent resistivity data to its corresponding true resistivity image. The proposed model consists of a CLIP-based text and image encoders to allow the mapping of both text and image pairs into a common embedding space. The Leaky Rectified Linear Unit activation function has shown better learning capabilities of the non-linear features of the data, making it suitable for feature extraction, while the Binary Cross-Entropy (BCE) loss function effectively discriminates between the real and synthesized resistivity images.

Finally, the proposed imaging model was able to achieve a Fréchet Inception Distance (FID) score and Structural Similarity Index (SSIM) of 265.10 and 0.62, respectively, in detecting different types of objects on the collected test data from a controlled small-scale resistivity imaging system. This resulted in a ~1% to 7% increase in quality and structural similarity compared to traditional inversion and existing state-of-the-art models. This study could potentially offer an introductory technique for fast generation of ground truth resistivity images, solving the need for significant computational time and resources.

Abstract Format

html

Language

English

Format

Electronic

Keywords

Image processing—Digital techniques; Deep learning (Machine learning)

Upload Full Text

wf_yes

Embargo Period

4-22-2025

This document is currently not available here.

Share

COinS