The Representation Learning Ability of Self-Supervised Learning in Unlabeled Image Data

Many existing systems struggle to strike a balance between global feature discrimination and local semantic understanding, despite the growing popularity of Self-Supervised Learning (SSL) for representation learning with unlabeled image data. This study introduces a novel SSL framework—Contrastive and Contextual Self-Supervised Representation Learning (C2SRL)—which integrates contrastive learning mechanisms with auxiliary context-based pretext tasks, specifically rotation prediction and jigsaw puzzle solving. The proposed C2SRL enhances two leading constructive models, SimCLR and MoCo, by incorporating contextual modules and a unified multi-task loss function, thereby improving the robustness and generalizability of the learned representations. A lightweight ResNet backbone is employed for encoding, followed by a dual-view augmentation strategy and a projection head that maps features into a contrastive embedding space. The proposed C2SRL outperforms existing SSL approaches in terms of classification accuracy and clustering coherence on the STL-10 and CIFAR-10 datasets, two benchmark datasets. It demonstrates strong scalability, as evidenced by its 89.6% mAP and 0.81 NMI, achieved using only 10% labeled data for fine-tuning. These results highlight the potential of combining contextual and contrastive learning objectives to generate rich, transferable visual representations for low-label or label-free applications.