Scholarlink

<p>Neurosymbolic artificial intelligence is a promising type of AI that integrates neural networks and symbolic knowledge to improve the performance, interpretability, and reliability of machine learning models and AI applications. We present a comprehensive examination of over 60 state-of-the-art methodologies published from 2017 and 2025 on the application of neurosymbolic techniques to computer vision (CV) tasks, including classification, detection, segmentation, and multimodal vision-language reasoning. We categorize them into methods that use external solvers, approximate methods, probabilistic methods, graph methods, and methods that incorporate logical constraints. In addition, we highlight the main frameworks of neurosymbolic AI for CV tasks and detail the image datasets from the computer vision domain typically used in applications, along with the changes applied to address neurosymbolic tasks in computer vision. The main contributions of our work are: <i>(i)</i> Investigating recent neurosymbolic AI techniques, with a focus on core strategies for integrated visual learning and reasoning; <i>(ii)</i> Identifying current state-of-the-art methodological improvements, and <i>(iii)</i> Highlighting high-impact research avenues for neurosymbolic AI in computer vision. By examining current gaps, the paper further outlines the necessary advances to deploy neurosymbolic AI models within the computer vision domain.</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

Neurosymbolic computer vision: a survey and perspective

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月6日

查看全文
  • Marcio Nicolau,
  • Claudio R. Jung,
  • Luis C. Lamb

摘要

Neurosymbolic artificial intelligence is a promising type of AI that integrates neural networks and symbolic knowledge to improve the performance, interpretability, and reliability of machine learning models and AI applications. We present a comprehensive examination of over 60 state-of-the-art methodologies published from 2017 and 2025 on the application of neurosymbolic techniques to computer vision (CV) tasks, including classification, detection, segmentation, and multimodal vision-language reasoning. We categorize them into methods that use external solvers, approximate methods, probabilistic methods, graph methods, and methods that incorporate logical constraints. In addition, we highlight the main frameworks of neurosymbolic AI for CV tasks and detail the image datasets from the computer vision domain typically used in applications, along with the changes applied to address neurosymbolic tasks in computer vision. The main contributions of our work are: (i) Investigating recent neurosymbolic AI techniques, with a focus on core strategies for integrated visual learning and reasoning; (ii) Identifying current state-of-the-art methodological improvements, and (iii) Highlighting high-impact research avenues for neurosymbolic AI in computer vision. By examining current gaps, the paper further outlines the necessary advances to deploy neurosymbolic AI models within the computer vision domain.