본문 바로가기
  • Home

Analysis of Transfer Learning Factors in CNN-Based Image Classification Performance

  • Journal of Internet of Things and Convergence
  • Abbr : JKIOTS
  • 2025, 11(6), 19
  • Publisher : The Korea Internet of Things Society
  • Research Area : Engineering > Computer Science > Internet Information Processing
  • Received : November 28, 2025
  • Accepted : December 17, 2025
  • Published : December 31, 2025

Woo Jung Ahn 1 In Seop NA 2

1전남대 데이터사이언스대학원
2전남대학교

Accredited

ABSTRACT

This study experimentally analyzed the effects of domain similarity, backbone complexity, and transfer strategy on image classification performance in CNN-based transfer learning. After pre-training on CIFAR-10 as the source domain, a 2×2×2 factorial design experiment was conducted applying Feature Extraction and Fine-Tuning strategies using VGG16 and InceptionV3 backbones on STL-10 (similar domain) and Fashion-MNIST (dissimilar domain). The experimental results showed that the combination of InceptionV3 and Fine-Tuning achieved the highest performance with 90.5% accuracy in the similar domain, while the combination of VGG16 and Feature Extraction maintained stable performance of 78.6% in the dissimilar domain. This study demonstrated that transfer learning performance is determined by the interaction among three factors rather than a single factor, and provides practical guidelines for selecting strategies in domain adaptation and small-scale data transfer learning design.

Citation status

* References for papers published after 2024 are currently being built.