Scholarlink

<p>Automated defect detection in automotive bushing components requires high accuracy for dimensional irregularities, low annotation costs, and efficient edge deployment. Height-information imaging captures variations that 2D images miss, yet existing methods require full annotation and produce heavy models. Vision-language models reduce annotation through pseudo-labeling but lack height-understanding and demand excessive computational resources. Here, we present HeightVL-Distill, a four-stage framework that integrates height-information imaging with vision-language capabilities and progressive knowledge distillation. The framework comprises Height-VLM Alignment that teaches models to interpret stratified height representations, Cross-Modal Height-Visual Uncertainty for strategic sample selection, Progressive Height-Aware Distillation that transfers knowledge to lightweight students, and a Height-Preserving Lightweight Architecture. Experiments on 20,000 bushing images demonstrate 97.3% mean Average Precision at IoU 0.5 (mAP@50) with only 12.8% annotation. The distilled model contains 42M parameters and processes 67 frames per second; this represents 95<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation> parameter reduction over the vision-language teacher.</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

Height-information guided vision-language distillation for annotation-efficient bushing defect detection

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Songquan Xiong,
  • Shikun Chen,
  • Yangguang Liu,
  • Junjie Mao,
  • Ruikai Qiu

摘要

Automated defect detection in automotive bushing components requires high accuracy for dimensional irregularities, low annotation costs, and efficient edge deployment. Height-information imaging captures variations that 2D images miss, yet existing methods require full annotation and produce heavy models. Vision-language models reduce annotation through pseudo-labeling but lack height-understanding and demand excessive computational resources. Here, we present HeightVL-Distill, a four-stage framework that integrates height-information imaging with vision-language capabilities and progressive knowledge distillation. The framework comprises Height-VLM Alignment that teaches models to interpret stratified height representations, Cross-Modal Height-Visual Uncertainty for strategic sample selection, Progressive Height-Aware Distillation that transfers knowledge to lightweight students, and a Height-Preserving Lightweight Architecture. Experiments on 20,000 bushing images demonstrate 97.3% mean Average Precision at IoU 0.5 (mAP@50) with only 12.8% annotation. The distilled model contains 42M parameters and processes 67 frames per second; this represents 95 \(\times\) parameter reduction over the vision-language teacher.