University of the Aegean
Deep learning and machine learning techniques for image classification and analysis
Abstract
dc:descriptionDeep learning and machine learning have profoundly transformed image classification and analysis, enabling major advances in medical diagnostics, environmental and industrial monitoring, real-time object recognition, and natural language understanding. Convolutional neural net- works (CNNs), long short-term memory (LSTM) networks, attention-based modules, and hybrid architectures now achieve unprecedented capability to extract meaningful representations from high-dimensional, multimodal data. Yet, despite these achievements, significant challenges remain in attaining optimal accuracy, strong out-of-sample generalization, interpretability, and computational efficiency, highlighting the need for systematic research that spans multiple do- mains.This doctoral thesis consolidates and extends the findings of several peer-reviewed papers to propose, implement, and evaluate novel deep learning and machine learning techniques for diverse visual and text–image applications. The research is driven by three overarching objectives: (i) to design and optimize CNN-based and hybrid models that integrate LSTMs, attention mechanisms, and transformer-style modules; (ii) to ensure robustness and scalability through data-centric strategies such as transfer learning, advanced augmentation, and domain adaptation; and (iii) to embed interpretability and explainability so that AI systems become transparent and clinically or operationally trustworthy. Extensive experimental investigations have been performed on real-world datasets covering a wide spectrum of applications. In the medical domain, these include histopathological breast cancer slides, white blood cell images, MRI scans for brain tumor and Alzheimer’s diagnosis, chest X-rays for pneumonia and COVID-19, ultrasound images of thyroid nodules, and mi- croscopic images of malaria parasites. In environmental and industrial monitoring, the thesis addresses high-resolution satellite scene classification, smart-waste detection, weather image recognition, and maritime oil-spill support. In object detection and general image recognition, it covers chess piece identification, sign language recognition, sports and natural scene classification, and real-time face mask detection. Natural language and text-based experiments extend the approach to sentiment and emotion analysis on large-scale social media and movie-review datasets, including multimodal fusions of text and images. Across these application areas, the proposed methods consistently surpass baseline tech- niques and many state-of-the-art alternatives. Rigorous evaluation—using accuracy, precision, recall, F1-score, AUC, and, when appropriate, the Matthews correlation coefficient—demonstrates not only high predictive accuracy but also robust generalization to unseen conditions, efficient training times, and manageable computational cost. The studies further show how careful ablation, parameter tuning, and cross-domain testing consolidate these gains into reproducible and scalable solutions.A key scientific contribution of the thesis is the identification of cross-cutting architectural and methodological principles that unify its diverse experiments. Deep yet well-regularized CNN cores, augmented when needed by attention or recurrent/transformer modules, emerge as a consistent design backbone. Data-centric strategies—comprehensive augmentation, transfer learning from large natural-image corpora, and domain adaptation across institutions, sensors, or linguistic platforms—prove indispensable for ensuring robustness in the face of noise, small sample sizes, and distribution shifts. Explainability is treated as an intrinsic requirement: saliency maps, Grad-CAM heatmaps, and model-agnostic tools such as SHAP and LIME reveal the visual or textual cues driving predictions, enabling experts to verify and refine the models. These principles, repeatedly validated in the Discussion chapter, constitute a unifying methodological framework that transcends individual applications.Beyond technical innovation, the thesis addresses the ethical and societal dimensions of artificial intelligence. Particular attention is paid to transparency, fairness, and sustainability—critical factors for regulatory compliance and for public acceptance in sensitive fields such as healthcare and environmental management. By demonstrating that high accuracy can coexist with interpretability, energy efficiency, and responsible data governance, the work contributes to the development of trustworthy AI systems aligned with current European and international guidelines.The dissertation is organized in four parts. The first introduces the research problem, for- mulates objectives, and surveys the relevant literature, identifying gaps that motivate the new contributions. The second part presents the proposed framework, detailing architectural design, mathematical foundations, and training strategies. The third part reports extensive experi- mental evaluations across all studied domains, supported by ablation analyses and comparative studies with state-of-the-art methods. The fourth part offers a comprehensive discussion of find- ings, distills cross-domain design principles, and concludes with limitations and well-grounded directions for future research. In summary, this thesis provides a comprehensive, experimentally validated study of deep learning and machine learning for image classification and analysis. By advancing unified and interpretable architectures, robust data strategies, and scalable deployment practices, it shows how artificial intelligence can achieve scientific rigor, real-world reliability, and societal respon- sibility. The integrated results demonstrate that carefully engineered AI models can address complex classification challenges in medicine, environmental and industrial monitoring, real- time object recognition, and natural language processing, reinforcing the transformative impact of deep learning on modern computational intelligence.
Degree
thesis:*- Grantor dc:publisher
- University of the Aegean
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kanavos, Athanasios
- Καναβός, Αθανάσιος
Subjects
dc:subject × 12- Βαθιά μάθηση
- Μηχανική μάθηση
- Ανίχνευση αντικειμένων
- Deep learning
- Machine learning
- Object detection
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/61049