Bibliometric Analysis of Few-Shot Deep Learning Techniques in Visual Recognition Tasks

This work intends to deliver a thorough bibliometric analysis of few-shot learning methodologies in visual recognition tasks, examining the growth, trends, and prospective directions of this swiftly advancing domain. The study aims encompass the identification of primary themes, methodologies, and application domains, alongside the revelation of probable research deficiencies and nascent trends. Bibliometric data was obtained from the Web of Science Core Collection (WOS-CC) database, concentrating on articles from 2017 to 2024. The Biblioshiny tool facilitated the examination of keyword correlations and the temporal progression of research themes through cooccurrence analysis and thematic evolution investigation. The analysis revealed significant increase in few-shot learning research for visual recognition tasks, evidenced by an annual growth rate of 54.49% and an increasing number of citations per document. Keyword co-occurrence analysis identified four primary clusters: visual recognition tasks, deep learning architectures, learning methodologies, and data preparation strategies. In certain shot learning for image identification, convolutional neural networks and various learning methodologies, including meta-learning, transfer learning, and metric learning, emerged as fundamental components. Thematic evolution research underscored the sustained importance of key subjects, as well as the rise of new study paths like feature fusion. The results distinctly shed light on the current state-of-the-art. They also illuminate potential revolutionary paths in few-shot learning for visual recognition. The research greatly influences the thorough development of more efficient and effective few-shot learning algorithms, with many potential implications for multiple applications, including surveillance, medical image processing, and autonomous systems.