Self-Supervised Visual Representation Learning for 2D Images: A Review
ITM Web of Conferences · 2026 · 93 · 02004
DOI: 10.1051/itmconf/20269302004 ↗Kurz gefasst
Fachartikel in der Zeitschrift ITM Web of Conferences, online erschienen am 13.11.2026. Verfasst von Ranzhe Zhang. Die Arbeit nennt 13 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Self-supervised learning (SSL) has emerged as the dominant paradigm for training visual backbones in modern computer vision, fundamentally transforming the field by eliminating the prohibitive labeling costs that have long limited supervised learning’s scalability. This survey systematically examines the four major SSL paradigms—contrastive instance discrimination, non-contrastive self-distillation, masked image modeling, and joint-embedding predictive architectures—with a critical focus on the often-overlooked trade-offs that underpin their performance rather than merely reporting benchmark results. The paper analyzes how aggressive augmentations simultaneously empower and restrict contrastive methods, why non-contrastive learning’s collapse prevention remains theoretically incomplete, why pixel reconstruction may be a flawed proxy for semantic understanding, and why recent predictive architectures offer conceptual elegance but limited empirical gains. Special emphasis is placed on the critical misalignment between standard evaluation protocols and real-world downstream utility, particularly the significant performance discrepancies between linear probing and fine-tuning that have systematically biased the field’s research direction and self-assessment.
Zitieren
APA
Zhang, R. (2026). Self-Supervised Visual Representation Learning for 2D Images: A Review. ITM Web of Conferences, 93, 02004. https://doi.org/10.1051/itmconf/20269302004
Vancouver
Zhang R. Self-Supervised Visual Representation Learning for 2D Images: A Review. ITM Web of Conferences. 2026;93:02004. doi:10.1051/itmconf/20269302004
BibTeX
@article{Zhang2026,
author = {Zhang, Ranzhe},
title = {Self-Supervised Visual Representation Learning for 2D Images: A Review},
journal = {ITM Web of Conferences},
year = {2026},
volume = {93},
pages = {02004},
doi = {10.1051/itmconf/20269302004}
}
EUVIMED