Transformers in computer vision: a PRISMA-based systematic review, taxonomy and future perspectives

Transformers reshaped computer vision by enabling global modeling of visual relationships through self-attention. Their rapid adoption across diverse vision tasks has created a need for an up-to-date systematic review. Following PRISMA 2020, we searched IEEE Xplore, ACM Digital Library, Scopus, Web of Science, and Google Scholar for English-language studies published from 2017 to June 10, 2026. We included accessible studies that applied transformer or attention-based methods to computer vision and met predefined publication and reporting-quality. Two reviewers independently screened the records, and reporting quality was assessed using a five-item checklist. Because the studies differed in tasks, datasets, metrics, and experimental settings, we compared their findings through narrative synthesis rather than meta-analysis. Of 2,721 records identified, 163 studies across nine computer vision domains were included. We examined architectural evolution and pre-training strategies and compared transformers with CNNs in terms of performance, efficiency, and robustness. The findings show a progression toward hierarchical, efficient, self-supervised, multimodal, and foundation-model architectures, although CNNs remain competitive in data-limited and resource-constrained settings. Unlike earlier surveys, this review combines a PRISMA-guided methodology with broad coverage of recent advances and proposes a dual-axis taxonomy based on architecture and application domain. The conclusions are limited by experimental differences, restriction to English-language studies, citation-based preprint selection, and restricted Google Scholar screening. This review provides a systematic, time-bounded reference for major vision transformer paradigms and applications. It was not registered and received no external funding; the completed PRISMA 2020 checklist is available at https://figshare.com/s/fbd2cb08dbf2a8d943f6 .