Yıldız U., Yazıcıoğlu İ. K., Küçük K., Akbulut O.
APPLIED SCIENCES, cilt.16, sa.19, ss.1-16, 2026 (Scopus)
-
Yayın Türü:
Makale / Tam Makale
-
Cilt numarası:
16
Sayı:
19
-
Basım Tarihi:
2026
-
Doi Numarası:
10.3390/app16199456
-
Dergi Adı:
APPLIED SCIENCES
-
Derginin Tarandığı İndeksler:
Applied Science & Technology Source, Scopus, Compendex, INSPEC, Directory of Open Access Journals
-
Sayfa Sayıları:
ss.1-16
-
Kocaeli Üniversitesi Adresli:
Evet
Özet
The quality of support sets plays a critical role in few-shot learning. How well class prototypes represent their corresponding categories depends directly on the samples selected to construct these sets during episodic training. In standard implementations of Prototypical Networks, support samples are typically selected at random, which may result in support sets composed of highly similar or redundant samples in the feature space, particularly under limited-data scenarios. Such redundancy compromises prototype quality and may degrade classification performance. To address this issue, this study proposes a Correlation-Aware Support Resampling (CASR) mechanism. CASR evaluates the feature-space distributions of candidate support sets using an independent MobileNetV2 feature extractor and identifies the most suitable support configuration through the mean off-diagonal correlation ρ¯ metric derived from the covariance matrix, aiming to reduce redundancy among support samples while preserving within-class representativeness. Specifically, CASR operates in three steps: (i) candidate support sets are randomly sampled from the available class pool; (ii) feature vectors are extracted using a frozen MobileNetV2, and the mean off-diagonal correlation ρ¯ is computed from the resulting covariance matrix; and (iii) the first configuration satisfying ρ¯>τ is accepted as the support set, or the initial random set is retained if no valid configuration is found within Tmax attempts. The proposed approach is designed as a lightweight and architecture-independent support-set refinement mechanism that operates without modifying either the internal architecture or the learning procedure of Prototypical Networks. Experiments conducted on the MiniImageNet and Omniglot datasets demonstrate that CASR consistently improves performance compared with standard Prototypical Networks based on random support sampling. On MiniImageNet under five-way, five-shot settings, CASR consistently outperforms the ProtoNet baseline across training episodes. At 1000 episodes, CASR achieves 62.28% with τ=0.25 compared to 61.62% for the baseline. At 1750 episodes, CASR achieves 66.2% with τ=0.25—the recommended setting offering the best trade-off between accuracy and computational cost—compared to 65.45% for the baseline, while τ=0.28 reaches 66.5% at 1750 episodes at the cost of approximately 14.6× higher training time. On Omniglot, where intra-class variance is inherently low, CASR achieves 99.56% at 1000 episodes and 99.66% at 1500 episodes with τ=0.30, compared to 99.54% and 99.60% for the baseline, respectively. These results confirm that the proposed mechanism is most effective in visually complex datasets with high intra-class variation.