Thumbnail Image
Publication

Advancing graph neural networks : addressing representation challenges in the vision domain

Citations
Altmetric:
Date
2025-12
Abstract
Graph Neural Networks (GNNs) offer a potent approach for capturing complex spatial and relational structures within images, thereby complementing the local feature extraction capabilities of Convolutional Neural Networks (CNNs). This is particularly advantageous for medical imaging and intricate computer vision tasks where explicit relationships are paramount. However, the effective application of GNNs to visual data encounters significant hurdles, including the generation of meaningful graph representations, managing prevalent low homophily, mitigating over-smoothing in deeper network layers, and preserving information during graph pooling operations. This PhD thesis systematically addresses these core challenges. Following a comprehensive review of graph-based methods in computer vision, which identified critical limitations, this research introduces novel GNN architectures. Specifically, a Graph Convolution Neural Network Enhanced Connectivity (GCNN-EC) is proposed to counteract over-smoothing by exploiting feature interrelationships. Furthermore, another innovative architecture is developed to specifically tackle low homophily through dynamic feature grouping and learned aggregation strategies. The research also explores hybrid architectures that effectively integrate CNNs and GNNs. This includes the development of GNN enhanced models for semantic segmentation across diverse image modalities, such as medical and distorted imagery, and the integration of graph-based techniques for efficient medical image analysis, particularly under data-constrained scenarios. Collectively, this thesis advances the state-of-the-art in applying GNNs to image understanding. By providing novel architectural solutions and integration strategies that overcome fundamental representational challenges, this work contributes more robust and effective models for classification and segmentation tasks in both general and medical vision domains.
Supervisor
Description
Peer-reviewed
Publisher
University of Limerick
Citation
Collections
Doctoral
Electronic & Computer Engineering
Show all metadata
Files
Thumbnail Image
Singh_2025_Advancing.pdf
Adobe PDF, 24.09 MB
Keywords
Graph Neural Networks, Convolutional Neural Networks, medical imaging, computer vision tasks
ULRR Identifiers
https://hdl.handle.net/10344/32020
 https://doi.org/10.34961/19623
Funding code
Funding Information
Sustainable Development Goals
(9) Industry, Innovation and Infrastructure
External Link
License
Attribution-NonCommercial-ShareAlike 4.0 International
cbn
Embedded videos