Scholarlink

<p>To address the challenges of large-scale variations, weak textures, and parameter redundancy in steel surface defect detection, a real-time lightweight hybrid multi-scale adaptive feature learning network, termed HMA-DETR, is proposed based on RT-DETR. First, a hybrid gated aggregation network (HGANet) is designed to enhance defect-related spatial–channel representation through multi-scale spatial modeling and lightweight channel refinement, thereby improving the representation of defects with diverse shapes and textures. Second, a triple re-parameterized reconstruction stack (TriRepStack) is introduced after multi-scale feature fusion to reconstruct fused features and reduce semantic discrepancies across different feature levels, while preserving inference efficiency through structural re-parameterization. Finally, an adaptive sampling convolution (ASConv) is introduced to perform learnable offset-guided sampling and feature rearrangement during spatial reduction, aiming to alleviate information loss for small and irregular defects. Experimental results on the GC10-DET and NEU-DET datasets show that HMA-DETR improves mAP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mn>50</mn> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation> by 3.8% and 1.6% over the RT-DETR-r18 baseline, respectively. Meanwhile, HMA-DETR reduces the parameter count and computational cost by 28.5% and 10.1%, respectively, and achieves 122.9 FPS on GC10-DET and 124.1 FPS on NEU-DET under the tested GPU inference setting. These results indicate that HMA-DETR achieves a favorable accuracy–efficiency trade-off for real-time steel surface defect detection.</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

HMA-DETR: a real-time hybrid multi-scale adaptive feature learning network for steel surface defect detection

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Jinxia Yu,
  • Yang Chen,
  • Feiyang Ma,
  • Yuhang Sun

摘要

To address the challenges of large-scale variations, weak textures, and parameter redundancy in steel surface defect detection, a real-time lightweight hybrid multi-scale adaptive feature learning network, termed HMA-DETR, is proposed based on RT-DETR. First, a hybrid gated aggregation network (HGANet) is designed to enhance defect-related spatial–channel representation through multi-scale spatial modeling and lightweight channel refinement, thereby improving the representation of defects with diverse shapes and textures. Second, a triple re-parameterized reconstruction stack (TriRepStack) is introduced after multi-scale feature fusion to reconstruct fused features and reduce semantic discrepancies across different feature levels, while preserving inference efficiency through structural re-parameterization. Finally, an adaptive sampling convolution (ASConv) is introduced to perform learnable offset-guided sampling and feature rearrangement during spatial reduction, aiming to alleviate information loss for small and irregular defects. Experimental results on the GC10-DET and NEU-DET datasets show that HMA-DETR improves mAP \(_{50}\) 50 by 3.8% and 1.6% over the RT-DETR-r18 baseline, respectively. Meanwhile, HMA-DETR reduces the parameter count and computational cost by 28.5% and 10.1%, respectively, and achieves 122.9 FPS on GC10-DET and 124.1 FPS on NEU-DET under the tested GPU inference setting. These results indicate that HMA-DETR achieves a favorable accuracy–efficiency trade-off for real-time steel surface defect detection.