Scholarlink

<p>Steel surface defect detection constitutes a critical process in modern manufacturing systems, where efficient inspection facilitates prompt remediation of non-compliant products to ensure structural integrity and operational safety. Image-based inspection methodologies encounter significant technical challenges including indistinct defect signatures; substantial morphological heterogeneity and computational inefficiency. Conventional computer vision approaches demonstrate inherent limitations including elevated false-negative rates, suboptimal feature integration, and prohibitive model redundancy. This research introduces “Li-YOLO Net” – an optimized architecture derived from YOLOv11 foundations – designed for enhanced surface defect detection. An Edge-Enhanced Feature Extraction module is integrated within the backbone network, synergistically combined with Deformable Attention mechanisms to establish a novel feature extraction framework for discriminative feature representation. A Multi-scale Aware Feature Pyramid Network (MS-FPN) facilitates cross-scale feature interaction through dedicated selection modules, resolving multi-level fusion deficiencies. A lightweight detection architecture employing parameter-shared convolutional schemes is developed, establishing a dynamic task-collaborative paradigm to eliminate prediction inconsistencies stemming from task decoupling. Comprehensive evaluations on NEU-DET and GC10-DET benchmarks demonstrate significant MAP improvements (NEU-DET: 0.766; GC10-DET: 0.717). While preserving real-time performance, the proposed framework achieves 36.8% parameter reduction and 20.6% computational complexity decrease, outperforming state-of-the-art models and delivering an industrially viable lightweight inspection solution.</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

Li-YOLO Net: a lightweight steel defect detection framework with dynamic feature selection and task alignment

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Hao Wang,
  • Hongfei Zhan

摘要

Steel surface defect detection constitutes a critical process in modern manufacturing systems, where efficient inspection facilitates prompt remediation of non-compliant products to ensure structural integrity and operational safety. Image-based inspection methodologies encounter significant technical challenges including indistinct defect signatures; substantial morphological heterogeneity and computational inefficiency. Conventional computer vision approaches demonstrate inherent limitations including elevated false-negative rates, suboptimal feature integration, and prohibitive model redundancy. This research introduces “Li-YOLO Net” – an optimized architecture derived from YOLOv11 foundations – designed for enhanced surface defect detection. An Edge-Enhanced Feature Extraction module is integrated within the backbone network, synergistically combined with Deformable Attention mechanisms to establish a novel feature extraction framework for discriminative feature representation. A Multi-scale Aware Feature Pyramid Network (MS-FPN) facilitates cross-scale feature interaction through dedicated selection modules, resolving multi-level fusion deficiencies. A lightweight detection architecture employing parameter-shared convolutional schemes is developed, establishing a dynamic task-collaborative paradigm to eliminate prediction inconsistencies stemming from task decoupling. Comprehensive evaluations on NEU-DET and GC10-DET benchmarks demonstrate significant MAP improvements (NEU-DET: 0.766; GC10-DET: 0.717). While preserving real-time performance, the proposed framework achieves 36.8% parameter reduction and 20.6% computational complexity decrease, outperforming state-of-the-art models and delivering an industrially viable lightweight inspection solution.