Back to search

University of Thessaly (UTH)

Color image analysis and optimization for shadow removal and segmentation

Abstract

dc:description

This dissertation aims at the investigation of color image processing and analysis methodologies focusing on the segmentation of regions of interest, shadow detection and removal. The segmentation regions of interest have various applications in medical imaging, such as the accurate delineation of possibly abnormal regions for measurement and assessment. The presence of shadows in natural environment images can affect intelligent object recognition and detection systems, as it alters the luminosity and features of objects and their environment. The aim is to develop methodologies to tackle these problems and to improve accuracy in relation to existing methods, to improve processing times, as well as to reduce the amount of data required for their adaptation in different applications. Nature-inspired optimization algorithms can contribute to these processes through the physical and mathematical approaches which are based on. Initially, the investigation of such algorithms led to the development of a novel algorithm, called Emotion-Aware Brain Storm Optimization (EMABSO), which is inspired by the emotional state of the participants during the brainstorming process of a human population. This algorithm has shown that it generally excels in solving optimization problems, based on an evaluation carried out using benchmark functions that have been proposed for this purpose in the literature. Then, in order with shadows, the original H-V Shadow Detection (HSVD) methodology was developed using an optimization algorithm inspired by the electromagnetic forces between charged particles. HVSD performed better than, or comparable to, state-of-the-art methodologies that either a) make use of expensive sensors in the near infrared region (NIR) of the electromagnetic spectrum, or b) use computationally expensive methods based on machine learning. The HVSD methodology was extended by developing the unsupervised Simple Shadow Removal Algoritghm (SUShe) shadow removal algorithm whose main advantages are a) that it does not require training on a large amount of data and b) its low computational complexity. The results showed that SUShe outperforms all unsupervised methods in the existing literature, as well as some supervised methods based on more complex techniques of Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs). Compared to those SUShe shows lower performance, it gives comparable results. In addition, aiming to investigate the removal of shadows in more complex images, an original Multiscale Shadow Removal U-Net (MSSRU-Net) architecture was developed, and compared with various state-of-the-art relevant methodologies. The proposed, and the state-of-the-art methodologies were evaluated on a dataset containing images that simulate real situations from everyday life, including real objects and obstacles in complex scenes. The results showed that MSSRU-Net outperforms the state-of-the-art methods. Finally, in the context of medical image analysis, a hybrid Vision Transformers (ViTs) training methodology was developed, which is based on a traditional segmentation technique based on level sets. The energy terms contained in these sets provide spatial information about the boundaries of the contours of the anomalies. Therefore, ViTs were integrated with the equilibrium sets by introducing a loss function inspired by them, during the training of the models. The methodology was fully evaluated on three public datasets: two including images of skin lesions and one including histopathological images of colon9adenocarcinomas. The results showed that the use of the specific function a) improves the performance of innovative ViTs and b) contributes to strengthening the generalization ability of these models. The works presented in this thesis include methods aiming to solve image segmentation and shadow detection issues. These methods can be used as a basis for future research and applications, such as robotics, assisted navigation and telemedicine.

Degree

thesis:*
Grantor dc:publisher
University of Thessaly (UTH)
Year dc:date
2023

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Koutsiou, Dimitra-Christina
  • Κούτσιου, Δήμητρα-Χριστίνα

Subjects

dc:subject × 14

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/56148

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

Koutsiou, Dimitra-Christina; Κούτσιου, Δήμητρα-Χριστίνα. Color image analysis and optimization for shadow removal and segmentation. University of Thessaly (UTH), 2023. http://hdl.handle.net/10442/hedi/56148