The Visual Computer Компьютерные науки (информатика)
Evaluating architectural innovations in deep learning-based image semantic segmentation
Аннотация
Abstract Image semantic segmentation, a pivotal task in computer vision, involves partitioning images into meaningful regions for precise pixel-level classification. Although numerous review articles have summarized advances in deep learning-based segmentation, limited attention has been given to systematically examining how architectural innovations interact with the characteristics of benchmark datasets. To address this gap, this review systematically investigates deep learning-based semantic segmentation studies evaluated on four widely adopted benchmarks: Cityscapes, CamVid, PASCAL VOC 2012, and ADE20K. A structured literature search and screening process identified 374 eligible studies, from which representative high-performing models were selected for detailed analysis based on benchmark relevance, architectural novelty, methodological significance, and reported performance. For each study, architectural components, context modeling strategies, attention mechanisms, feature fusion approaches, segmentation performance, limitations, and future research directions were extracted and compared. The synthesis reveals that recent advances are increasingly driven by the integration of multi-scale context aggregation, attention-based feature refinement, transformer-based contextual modeling, hierarchical feature fusion, self-supervised pretraining, and vision-language alignment. Furthermore, the findings demonstrate that the effectiveness of architectural strategies is strongly influenced by dataset characteristics, including scene complexity, object scale variation, class granularity, temporal structure, and domain diversity. By establishing explicit links between dataset properties and architectural design choices, this review provides a dataset-centric taxonomy and a comprehensive synthesis of emerging design principles for developing robust and generalizable semantic segmentation models.