Back to search

University of Thessaly (UTH)

Intelligent systems and services for image and video analysis

Abstract

dc:description

This doctoral dissertation explores intelligent systems and services for image and video analysis. In view of scientific challenges for developing innovative solutions with a broad social impact, it investigates applications in biomedicine and computer-assisted navigation of visually impaired individuals. In this context, it focuses on machine learning, particularly the investigation of methods to improve the efficiency and the effectiveness of deep artificial neural network architectures, such as the Convolutional Neural Networks (CNNs).In Convolutional Neural Networks (CNNs) the input data can contain uncertainties, such as noise, color and geometric ubiquities, that is naturally propagated from the input layer to the convolution layers of the network affecting the quality of the extracted features. To cope with this problem, a novel pooling operation based on (type-1) fuzzy sets is proposed, named Fuzzy Pooling, which can be used as a drop-in replacement of the current, crisp, pooling layers of CNN architectures. Several experiments using publicly available datasets show that the proposed approach can enhance the classification performance of a CNN.Aiming to improve the effectiveness of CNNs, especially in the context of medical image analysis, a novel architecture named Look Behind Fully Convolutional Neural Network (LB-FCN) is proposed. The architecture is capable of extracting multi-scale image features by using blocks of parallel convolutional layers with different filter sizes. These blocks are connected by look-behind connections, so that the features they produce are combined with features extracted from behind layers, thus preserving the respective information. Furthermore, it has a smaller number of free parameters than conventional CNN architectures, which makes it suitable for training with smaller datasets. This is particularly useful in medical image analysis, since data availability is usually limited, due to ethicolegal constraints. Experiments on publicly available gastrointestinal image datasets show higher classification performance compared to state-of-the-art machine and deep learning methodologies. The architecture is capable of generalizing well even when the training dataset is different than the one on which it is tested. To investigate that, a novel cross-dataset experimental study was performed on various publicly available gastrointestinal tract image datasets, containing images from different modalities, including Wireless Capsule Endoscopy (WCE) and flexible endoscopy. The number of training samples in CNN training is directly linked to their generalization performance. When the training samples are limited, such as in the case of medical images, the generalization performance is negatively affected. A typical approach to mediate this problem is to use data augmentation techniques, which image rotation and translation. While effective, this technique still requires a substantial amount of training samples to be available. To battle this problem, in the context of inflammatory conditions detection in WCE images, a novel approach is presented that uses Generative Adversarial Networks (GANs) to generate artificial images. More specifically the study trained two GANs, one to generate healthy small bowel images and another, images with inflammatory conditions. The images are then used to train a CNN architecture and validate its performance in real images. The results from this study show that the substitution of real with artificially generated endoscopic images for CNN training can be a viable option.While CNNs have a remarkable performance in computer vision problems, usually, they are computationally expensive. This limits their usage in high-end expensive devices with multiple graphical processing units (GPUs). To mediate the problem, a typical approach is to reduce the number of floating-point operations (FLOPs) required for inference, at the expense of generalization performance. In this context, a novel LB-FCN inspired CNN architecture was proposed, named LB-FCN light. The architecture features a relatively low number of free parameters and FLOPs, while managing to maintain high generalization performance. The performance of the network is validated in the problem of staircase detection in indoor and outdoor environments, with application on assisted navigation of visually impaired individuals. The results from the experimental evaluation of LB-FCN light indicate its advantageous performance over the relevant state-of-the-art architectures.The development of easy-to-use machine learning (ML) application frameworks has enabled the development of advanced artificial intelligence (AI) applications with only a few lines of self-explanatory code. However, the deployment of ML algorithms as a service for remote high throughput ML task execution, involving complex data-processing pipelines can still be challenging, especially with respect to production ML use cases. To cope with this issue, a novel system architecture is presented, which enables Algorithm-agnostic, Scalable ML (ASML) task execution for high throughput applications. It aims to provide an answer to the research question of how to design and implement an abstraction framework, suitable for the deployment of end-to-end ML pipelines in a generic and standard way. The architecture manages horizontal scaling, task scheduling, reporting, monitoring and execution of multi-client ML tasks using modular, extensible components that abstract the execution details of the underlying algorithms. Applications of ASML are investigated for the analysis of image streams in the context of medical image analysis and assisted navigation of visually impaired individuals. The results of the experiments performed demonstrate its capacity for parallel, mission critical, task execution. Assistive navigation systems require the development, assessment, and optimization of different algorithms for obstacle detection, recognition, and avoidance, as well as path planning. This is a painstaking and costly process that requires repetitive measurements under stable conditions, which is usually difficult to achieve. To this end, a novel digital twin framework for the simulation and evaluation of assistive navigation systems is presented. The framework can replicate relevant real-life situations, enabling the evaluation and optimization of algorithms through adjustable and cost-effective simulations. The utility and the effectiveness of the framework are demonstrated with an indicative simulation study in the context of a camera-based wearable system for the navigation of visually impaired individuals in an outdoor cultural space.The work presented in this dissertation includes methods with both theoretical and practical impact, that can be used as the basis for further research, and the applications presented can be used as paradigms for applications on different domains, such as telemedicine, robotics, and intelligent transportation systems.

Degree

thesis:*
Grantor dc:publisher
University of Thessaly (UTH)
Year dc:date
2021

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Diamantis, Dimitrios
  • Διαμαντής, Δημήτριος

Subjects

dc:subject × 24

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/49853

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

Diamantis, Dimitrios; Διαμαντής, Δημήτριος. Intelligent systems and services for image and video analysis. University of Thessaly (UTH), 2021. http://hdl.handle.net/10442/hedi/49853