Back to search

Democritus University of Thrace (DUTH)

Optimization techniques for image fusion and image analysis

Abstract

dc:description

Image fusion and image analysis aim at the development of fused images that overcome limitations of camera lenses and sensors. Fused images should be more informative than all input images without introducing artifacts during fusion. However, the analysis of the input images and the development of fusion rules are an open research issue. Optimization techniques aim to analyze the input images and merge them in order to create a fused image. Optimization techniques for image fusion and image analysis have become important in the last decades.In order to perform image fusion, reliable machine learning techniques that lead to optimal image fusion solution must be developed. Recent research efforts on image fusion have revealed several weaknesses of current fusion methods. Most image fusion methods are prone to artifacts, when the input images are likely to have mis-registration or depict a dynamic scene. Moreover, since input images are likely to have noise due to various sources, such as camera sensors and image quantization, the development of robust-to-noise methods is of high importance. However, the issue of the noisy input images is not widely studied and most of fusion methods are likely to be sensitive to image noise.The present thesis is an effort to resolve the aforementioned problems and develop optimization techniques that assist in analyzing the input images and lead to fused images of high quality. In the first stage of this effort, a robust-to-Gaussian noise image fusion framework is proposed. In the second stage, an image fusion framework with robustness against Gaussian, Salt & Pepper and Poisson noise is developed. Finally, a fusion framework that is robust to Gaussian noise and provides image fusion and denoising during fusion is proposed.Image fusion is employed, in order to cope with the hardware limitations of camera sensors and camera lenses. In this work, three different methodologies to cope with the multi-focus image fusion problem with novel optimization techniques are introduced. Each of the proposed methods investigates the image fusion problem with clean input images and input images that contain different types of noise. The novelty of the proposed methods lies in the development of novel optimization techniques in order to analyze the input images and make the fusion frameworks robust against potential noise and mis-registrations found in the input images. In the first method, a Conditional Random Field model with hand-crafted unary and smoothness potentials is introduced. The method is robust against Gaussian noise and has applications to medical image fusion and visible-infrared image fusion. In the second method, Deep Learning Networks are trained to compute unary and smoothness potential functions of higher complexity and afterwards a Conditional Random Field model is developed to describe the fusion problem. The method is robust to Gaussian noise, Salt & Pepper noise and Poisson noise. In the last method, a Conditional Random Field model is used to guide the transform domain image fusion method. The proposed method, is robust against Gaussian noise and supports denoising during fusion. Experimental results demonstrate that all methods outperform state-of-the-art image fusion methods in perfectly registered clean input images, in input images with potential mis-registration and are more robust against noise.The contribution of the presented techniques is the development of optimization techniques to both analyze input images and perform image fusion of high visual quality. The proposed models analyze the input images to find well-focused pixels and discriminate them from the out-of-focus or blurred ones and use conditional random field models to find a global or close to global optimum fused image result. They provide fused images of high visual quality without introducing artifacts during fusion and can be useful tools for image fusion of multi-focus images, medical image fusion and visible-infrared image fusion.

Degree

thesis:*
Grantor dc:publisher
Democritus University of Thrace (DUTH)
Year dc:date
2022

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Bouzos, Odysseas
  • Μπούζος, Οδυσσέας

Subjects

dc:subject × 12

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/52627

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Bouzos, Odysseas; Μπούζος, Οδυσσέας. Optimization techniques for image fusion and image analysis. Democritus University of Thrace (DUTH), 2022. http://hdl.handle.net/10442/hedi/52627