Learning Semantic Consistency for Robust Image Segmentation

Abstract

Image segmentation models often fail when they encounter unseen lighting, motion blur, or domain shift between synthetic and real data. This dissertation proposes a unified framework—Semantic‑Consistency Learning (SCL)—to deliver stable, fine‑grained segmentation under such challenging conditions. First, a dual‑branch architecture aligns class‑specific features from weakly and fully supervised streams, allowing the network to learn from sparse or noisy labels while keeping annotation costs low. Next, a temporal self‑ensembling strategy tracks an exponential moving average of network parameters, smoothing gradients and increasing resistance to distribution shift. Third, frequency‑domain perturbations and adaptive contrastive losses are introduced to suppress artefacts caused by weather, night scenes, and motion. Comprehensive tests on Cityscapes, ACDC, SynROD, and custom UAV datasets show mean‑IoU gains of 3–5 % over strong baselines, with a 40 % reduction in required pixel annotations. Ablation studies confirm that each consistency constraint contributes additively to accuracy and generalisation. The thesis also provides theoretical bounds explaining why multi‑scale consistency regularises feature space and improves robustness. Finally, an open‑source 3‑D visualisation toolkit is released, enabling interactive scrutiny of uncertainty maps and facilitating downstream use in autonomous driving and low‑altitude logistics planning. Taken together, these contributions advance the practical deployment of segmentation models in safety‑critical and resource‑constrained scenarios.

Suggested keywords (≤ 6, comma‑separated)

semantic segmentation, consistency learning, domain adaptation, deep learning, robust vision, annotation efficiency