Scholarlink

Existing defect detection methods struggle with weak abnormal patterns in industrial imaging, often leading to blurred edges and low-contrast feature degradation. While vision-language models have been integrated into recent anomaly detection frameworks, they fail to capture microscopic glass defects like micro-cracks and bubble clusters. This paper introduces two main contributions: First, we present GlassDefect, an industrial glass anomaly dataset with 712 samples across 7 defect categories. Second, we propose LUMA, an LLM-driven Adaptive Processing Framework, consisting of: 1)ATR (Adaptive Text Refinement), where the LLM updates textual descriptions based on DINO-X localization results; 2)ACIS (Attribute-Constrained Instruction Synthesizer), where the LLM generates operational attributes and instructions based on visual data and command inputs; 3)IOIM (Instruction-Operation Interaction Module), which enhances features adaptively. LUMA dynamically generates instructions based on defect size and reflection intensity, reducing reliance on manual configuration and optimizing processing workflows. The framework achieves significant improvements on the MVTec-AD and GLASS-Defect datasets, demonstrating its potential for real-world industrial applications.

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

LUMA: An LLM-Driven Unified Multimodal Framework for Glass Detection via Adaptive Text Updates and Instruction Scheduling

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pinjie He,
  • Mingming Wang,
  • Xinpan Yuan,
  • Xianjun Li

摘要

Existing defect detection methods struggle with weak abnormal patterns in industrial imaging, often leading to blurred edges and low-contrast feature degradation. While vision-language models have been integrated into recent anomaly detection frameworks, they fail to capture microscopic glass defects like micro-cracks and bubble clusters. This paper introduces two main contributions: First, we present GlassDefect, an industrial glass anomaly dataset with 712 samples across 7 defect categories. Second, we propose LUMA, an LLM-driven Adaptive Processing Framework, consisting of: 1)ATR (Adaptive Text Refinement), where the LLM updates textual descriptions based on DINO-X localization results; 2)ACIS (Attribute-Constrained Instruction Synthesizer), where the LLM generates operational attributes and instructions based on visual data and command inputs; 3)IOIM (Instruction-Operation Interaction Module), which enhances features adaptively. LUMA dynamically generates instructions based on defect size and reflection intensity, reducing reliance on manual configuration and optimizing processing workflows. The framework achieves significant improvements on the MVTec-AD and GLASS-Defect datasets, demonstrating its potential for real-world industrial applications.