Scholarlink

<p>This study presents an integrated framework for multimodal metadata enrichment that combines image feature extraction with vision–language model (VLM) keyword generation to enhance the descriptive quality of metadata for AI-generated images. Conventional keyword-generation methods often inadequately represent the perceptual and aesthetic complexity inherent in AI-generated images. The proposed framework addresses this gap by integrating ten visual attributes, including color (name, aspect, scheme, effect, psychology, and temperature), balance, texture, contrast, and brightness, with VLM-derived keywords to form a coherent and reproducible metadata model. Using a dataset of 20,000 AI-generated images, two keyword sets were evaluated: a VLM-only baseline and a feature-enriched combined set. Reverse image generation was performed with Microsoft Copilot Designer under identical conditions, and performance was assessed using image-similarity metrics (Euclidean distance, cosine similarity, and histogram intersection) and keyword-accuracy metrics (recall, precision, and F1-score). The combined approach achieved statistically significant improvements: Euclidean distance decreased by 25.4%, cosine similarity increased by 25.4%, and histogram intersection improved by 33.8%. Correspondingly, recall, precision, and F1-score increased by 32.8%, 31.4%, and 22.2%, respectively (<i>p</i> &lt; .01). An ablation study confirmed complementary contributions across feature groups, with color-based descriptors yielding the highest visual-similarity gains and texture features contributing most to retrieval accuracy. Beyond empirical performance, the framework advances a theoretical model that unifies perceptual and linguistic modalities, supporting the development of explainable, interoperable, and scalable metadata systems for digital libraries, creative industries, and AI-driven content management. While the implementation currently depends on proprietary platforms (GPT-4o and Copilot Designer), the workflow itself is model-agnostic and reproducible using open-source alternatives. Future work will extend these evaluations with open models to enhance transparency and alignment with open-science principles.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improving AI-Generated image metadata: A detailed approach using image feature extraction

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Akara Thammastitkul

摘要

This study presents an integrated framework for multimodal metadata enrichment that combines image feature extraction with vision–language model (VLM) keyword generation to enhance the descriptive quality of metadata for AI-generated images. Conventional keyword-generation methods often inadequately represent the perceptual and aesthetic complexity inherent in AI-generated images. The proposed framework addresses this gap by integrating ten visual attributes, including color (name, aspect, scheme, effect, psychology, and temperature), balance, texture, contrast, and brightness, with VLM-derived keywords to form a coherent and reproducible metadata model. Using a dataset of 20,000 AI-generated images, two keyword sets were evaluated: a VLM-only baseline and a feature-enriched combined set. Reverse image generation was performed with Microsoft Copilot Designer under identical conditions, and performance was assessed using image-similarity metrics (Euclidean distance, cosine similarity, and histogram intersection) and keyword-accuracy metrics (recall, precision, and F1-score). The combined approach achieved statistically significant improvements: Euclidean distance decreased by 25.4%, cosine similarity increased by 25.4%, and histogram intersection improved by 33.8%. Correspondingly, recall, precision, and F1-score increased by 32.8%, 31.4%, and 22.2%, respectively (p < .01). An ablation study confirmed complementary contributions across feature groups, with color-based descriptors yielding the highest visual-similarity gains and texture features contributing most to retrieval accuracy. Beyond empirical performance, the framework advances a theoretical model that unifies perceptual and linguistic modalities, supporting the development of explainable, interoperable, and scalable metadata systems for digital libraries, creative industries, and AI-driven content management. While the implementation currently depends on proprietary platforms (GPT-4o and Copilot Designer), the workflow itself is model-agnostic and reproducible using open-source alternatives. Future work will extend these evaluations with open models to enhance transparency and alignment with open-science principles.