ADVANCEMENTS IN IMAGE SEGMENTATION: FROM SUPERVISED LEARNING TO ZERO-SHOT FOUNDATION MODELS

Abstract

Image segmentation is a fundamental task in computer vision with critical importance in biomedical imaging, environmental monitoring, and artificial intelligence. This dissertation investigates the evolution of segmentation methodologies across four major paradigms: traditional machine learning with handcrafted features, deep learning architectures, reinforcement learning–enhanced active learning, and zero-shot segmentation enabled by large vision–language models. Early work demonstrates that classical pixel-wise classifiers such as random forests and regression models, applied within the COLD framework, can achieve meaningful segmentation and land-change detection despite limited annotated data, though their reliance on handcrafted features restricts scalability. The transition to deep learning, exemplified by U-Net rectum segmentation on low-field MRI, improved accuracy by learning hierarchical features directly from data but remained constrained by small datasets, annotation burden, and limited cross-modality generalization. To reduce labeling costs and address dataset imbalance, this dissertation develops a reinforced active learning framework that integrates reinforcement learning with region-level uncertainty sampling, enabling efficient selection of informative samples and improving segmentation performance under constrained annotation budgets. Finally, the dissertation introduces a prototype-guided zero-shot segmentation framework that integrates bounding box priors from large vision–language models (Gemini Pro 2.5) with mask candidates from SAM and prototype-based similarity matching using CLIP embeddings. This training-free approach achieves robust segmentation across brain MRI, fetal ultrasound, and chest X-ray without task-specific fine tuning, reducing the performance gap between zero-shot and supervised models. Collectively, these contributions trace a technological progression from handcrafted feature engineering toward adaptable, annotation-efficient, and training-free segmentation paradigms. The findings highlight the growing potential of foundation models to democratize advanced image analysis in resource-limited environments. Future work includes improving LVLM-generated spatial priors, developing adaptive prototype selection strategies, and extending zero-shot frameworks to broader biomedical and environmental applications.