Home // International Journal On Advances in Software, volume 19, numbers 1 and 2, 2026 // View article
Authors:
Yurij Mikhalevich
Keywords: image tagging; zero-shot classification; CLIP; vision-language models; image metadata.
Abstract:
This paper presents an extended study of rtag, a practical, scalable, command-line tool for automatic semantic image tagging that leverages the Contrastive Language-Image Pre-Training model. Extending the original conference paper, this work contributes a survey of modern vision-language models and image tagging approaches, an in-depth analysis of prompt engineering strategies for zero-shot tagging, and a systematic comparison with existing tagging models and photo management applications. Evaluating three prompt templates on the ObjectNet dataset, the best template improves top-1 accuracy by 4.36 percentage points (from 25.37% to 29.73%) and top-5 accuracy by 5.94 percentage points (from 48.09% to 54.03%), substantially exceeding the improvement previously reported on ImageNet. We further discuss similarity threshold selection and its effect on tagging precision and recall, as well as standard image metadata formats and their role in integrating automated tagging into photo management workflows. The tool caches image features in an embedded database, enabling efficient re-tagging of previously indexed collections; the scalability analysis confirms approximately linear scaling of both indexing and tagging with the image count, with the cache providing a 25.3% speedup for previously indexed images. These results show that flexible, zero-shot image tagging with arbitrary user-defined vocabularies is practical at scale on consumer hardware, benefiting sectors that handle large volumes of visual content such as media and entertainment, digital archiving, healthcare, and personal photo management.
Pages: 13 to 27
Copyright: Copyright (c) to authors, 2026. Used with permission.
Publication date: June 30, 2026
Published in: journal
ISSN: 1942-2628