<![CDATA[Research evolution and trends in computer vision (CV) are important for understanding the field’s land-scape. Trends show which topics are gaining attention, while evolution reveals how those topics change over time. Understanding both helps researchers gain insight into CV and anticipate emerging areas of focus. How-ever, the rapid growth of publications makes such detection challenging. This paper aims to detect research evolution and trends in CV using topic modeling (TM) and large language model (LLM) techniques. The study applies TM and LLM approaches to papers from leading CV conferences, Computer Vision and Pattern Recog-nition (CVPR), International Conference on Computer Vision (ICCV), and Winter Conference on Applications of Computer Vision (WACV), published between 2013 and 2023, totaling more than 21,000 papers, using only abstracts and titles. The TM methods used are Latent Dirichlet Allocation (LDA) and Bidirectional Encoder Representations from Transformers for Topic Modeling (BERTopic), which generate keywords that represent topics. LLMs then refine these topics to support better analysis. The results show that research evolution and trends are easier to identify from abstracts than from titles, with BERTopic outperforming LDA in internal va-lidity based on coherence metrics and external validity based on human judgment. In addition, the topics evolved from traditional image processing tasks in earlier years to a stronger focus on deep learning and, more recently, generative approaches. Integrating TM techniques with LLMs enhances the detection of evolving re-search themes and trends in CV. This approach provides a clearer understanding of the field's development and helps anticipate future directions.]]>
Copyrights © 2026
