About

I am a Postdoctoral researcher in Deep Learning and Computer Vision at the University of Oxford, working with Prof Ben Sheldon.

My research investigates how inductive biases derived from visual perception and physical signal formation shape representation learning in deep neural networks. I focus on embedding principles such as frequency selectivity, texture shape decomposition, and modality aware priors into learnable architectures, with the aim of improving robustness, interpretability, and generalisation under real world conditions.

A central theme of my work is the design of novel convolutional and hybrid CNN–Transformer architectures that explicitly encode perceptual and physical priors. Rather than treating non-RGB sensing modalities as domain specific applications, I use them as stress tests for modern deep learning, exposing systematic failure modes in texture biased models and motivating principled architectural corrections.

I also study multimodal and multi-spectral learning, where I treat the physical properties of sensed data as a source of inductive bias rather than nuisance variability. By grounding model design in the physics of image formation, I develop data fusion strategies that improve cross-sensor generalisation and failure mode interpretability, particularly in visually complex and densely populated scenes.

My methodological contributions are validated on challenging real world datasets, including satellite and UAV imagery, where uncontrolled conditions such as occlusions, reflections, shadows, seasonal variation, and domain shifts are unavoidable. These settings provide a rigorous testbed for advancing representation learning beyond synthetic or overly curated benchmarks.

Current research themes:

Representation learning with explicit inductive biases, including frequency aware and perceptually grounded convolutional operators.
Semantic segmentation architectures featuring custom convolutional layers for complex visual domains.
Hybrid CNN–Transformer models that optimise the synergy between local feature extraction and global context modelling.
Multimodal and multi-spectral data fusion as a testbed for robust learning under physical constraints.
UAV based segmentation models resilient to occlusions, reflections, shadows, and illumination variability.
Domain-invariant feature learning for improved cross-sensor and cross-season generalisation.

Datasets and Experimental Infrastructure To support and stress test methodological development, I curate high quality datasets that expose real world failure modes in computer vision and remote sensing models. These datasets are designed not as standalone contributions, but as measurement instruments for analysing representation robustness, inductive bias, and generalisation behaviour under physically grounded variability.

Background:

I completed an EPSRC-funded PhD at the University of Sussex supervised by Prof. Andy Philippides and Prof. Novi Quadrianto, where my research focused on salient feature representation, domain adaptation, and semantic segmentation. This work laid the foundations of my current research agenda by investigating how feature representations degrade under domain shift and how architectural design choices influence generalisation.

As part of the Predictive Analytics Lab (PAL), I contributed to a British Academy-funded project on Satellite/Aerial Image Scene Segmentationon satellite and aerial image scene segmentation. In this context, I designed deep learning models for land use classification and multi-domain, multi-temporal data integration, using large scale real world imagery to study robustness between domains. Please see the demo below. demo During my research internship at Satellite Applications Catapult (supervised by Dr Cristian Rossi), I applied physics aware machine learning to industrial remote sensing, developing models that leveraged physical properties such as temperature and soil moisture to detect and characterise cement production facilities in China. This experience directly informed my subsequent work on embedding physical signal properties as inductive bias in deep learning architectures.

Industrial Experience:

Prior to my doctoral studies, I worked as a Control and Automation Design Engineer in the oil and gas industry. This role provided hands on experience with sensor driven systems, signal acquisition, control architectures, and safety critical software development, shaping my long standing interest in how physical processes, sensing constraints, and system design interact with data driven models.

Open to Collaboration:

I am always interested in discussing new research directions and potential collaborations, particularly around representation learning, inductive bias, and robust perception under real world sensing constraints. Feel free to get in touch!

Dr Georgios Voulgaris

Recent News