Democritus University of Thrace (DUTH)
Deep learning algorithms for semantic and instance segmentation
Abstract
dc:descriptionThis dissertation focuses on the image segmentation field through the study, development and application of Deep learning architectures on the most important and widely applicable field areas: semantic segmentation, document binarization, and instance segmentation. A strong emphasis is given on efficient, low complexity architectures that strike a balance between segmentation precision and computational efficiency, a combination that is crucial for practical real world applications. The work is structured into four chapters, one for the general image segmentation problem and one for each studied field area. Chapter 1 lays the groundwork for the dissertation by exploring image segmentation as a more generic research area and presenting its key subfields such as object detection, semantic and instance segmentation. Document binarization is also referenced as a distinctive semantic segmentation case. Objectives and challenges of each area are studied as well as the evaluation methods used each time. The chapter also touches upon image classification, since pretraining architectures on the classification problem is a foundational part of the overall segmentation pipeline. The purpose of the chapter is to stand as a theoretical and practical introduction, establishing the importance of image segmentation in modern automated real world applications. In chapter 2 the challenging problem of document binarization is examined, as a critical step in the transformation of a document, typed or handwritten, into the digital world. A retrospective look into the popular methods of the past takes place, emphasizing in the qualities that enabled them to govern the field for decades before the advent of Deep learning. In addition to that, existing state-of-the-art convolutional methods are also explored and finally a top performing convolutional architecture is presented that combines modern architectural traits, such as multi resolution blocks and visual attention mechanisms. The method is compared in terms of accuracy, robustness and speed with the state-of-the-art contemporary methods on the (H)DIBCO datasets that have become a standard document binarization benchmark. Chapter 3 presents a comprehensive study of semantic segmentation, a task whose goal is to assign class labels to all the pixels of an image. The chapter goes through a systematic review of the traditional semantic segmentation methods before delving into the Deep learning world, where both convolutional and transformer based methods are examined. The chapter also presents a novel, low complexity but highly competitive Deep learning architecture for semantic segmentation, the MResTNet, that combines both convolutional and transformer architectural blocks. Through a series of experiments and evaluations, the motivation behind the architecture is presented and its efficacy is explored. Chapter 4 examines the additional challenges of the instance segmentation subfield, where not only pixel level class labeling has to take place but also individual instance detection. Semantic segmentation and object detection principles have to be combined and work hand-in-hand so that individual object instances are labeled and distinguished, even in cases where they overlap each other. This chapter explores advanced Deep learning techniques, examining the advantages and flaws of each one. The work also presents a new approach for enhancing the precision and speed of instance segmentation models, maintaining high accuracy and yet keeping low, real-time complexity, the DDMnet. The method is analyzed and evaluated against the competition, on the dominant area benchmark datasets, showcasing its high efficiency and low computational complexity. Finally, chapter 5 summarizes the work in each of the previous research steps and describes the main focus and targets that guided the research activities. Additionally, it outlines the value and contributions of each suggested method to the field of image segmentation and concludes with a list of promising and innovative potential future research directions, inspired by the proposed architectures.
Degree
thesis:*- Grantor dc:publisher
- Democritus University of Thrace (DUTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Detsikas, Nikolaos
- Δέτσικας, Νικόλαος
Subjects
dc:subject × 16- Εικόνα, κατάτμηση
- Σημασιολογική, κατάτμηση
- Κατάτμηση, στιγμιότυπου
- Βαθιά, μάθηση
- Δυαδικοποίηση, εγγράφων
- Image, segmentation
- Semantic, segmentation
- Instance, segmentation
- Deep, learning
- Document, binarization
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Τεχνολογία μέσων
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Media Technology
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60008