Zaguán. Repositorio Digital de la Universidad de Zaragoza
Publicación
Artículo científico (JournalArticle)
Evaluation of transfer learning by means of fuzzy logic for hand gesture recognition
Hand gesture recognition (HGR) is increasingly adopted in real-world applications, yet selecting the most suitable model–hyperparameter configuration remains a time-consuming and opaque process. This work introduces a systematic evaluation framework for static HGR that combines transfer learning (TL) with a fuzzy logic expert system to provide a reproducible and transparent ranking of candidate models. We benchmark eight widely used CNN backbones across 64 fine-tuned configurations, analyzing their performance in terms of accuracy, recall, and F1-score, while explicitly documenting the fuzzy inference rules used to integrate these metrics. Experiments on 3-class and 5-class subsets of the Tiny Hand Gesture dataset consistently highlight the robustness of classical architectures such as VGG16 and ResNet50, with the best configuration achieving an F1-score of 0.995. Beyond accuracy, we extend the analysis to computational efficiency, showing that deeper or next-generation networks often incur higher training time and memory costs without yielding performance gains. Finally, we report an exploratory hybrid quantum–classical experiment based on ResNet50 with a variational quantum layer, which improved test accuracy from 0.876 to 0.94, at the expense of extremely high training costs. Overall, our contribution is threefold: (i) a reproducible TL evaluation pipeline with fuzzy ranking, (ii) an evidence-based comparison of classical versus modern CNN architectures in static HGR, and (iii) an initial exploration of quantum machine learning integration for gesture classification.