Scholarlink

<p>Steel plate surface defect detection remains challenged by the simultaneous demands for real-time processing, multi-scale defect identification, and robustness against complex backgrounds. Addressing these demands constitutes a computationally intensive task that inherently requires high-performance computing (HPC) capabilities, including graphics processing unit (GPU)-accelerated parallel training and heterogeneous neural processing unit (NPU)-based real-time edge inference. To address these challenges, this study proposes an improved, lightweight DETR with improved matching (DEIM) framework that systematically integrates state space model (SSM)-based global modeling, wavelet-domain edge enhancement, hierarchical cross-spatial recalibration, and Fourier-domain defect-background separation. Specifically, a lightweight SSM-based backbone is introduced to achieve global receptive field coverage with linear computational complexity via recursive inference. It is combined with wavelet transform and multi-kernel depthwise convolution to enhance high-frequency edge extraction and multi-scale feature representation. A hierarchical feature fusion encoding (HFFE) structure is then designed to bridge semantic gaps between different levels and suppress background texture interference through cross-spatial weight recalibration and coordinate attention. Furthermore, a spatial-frequency dual-domain feature enhancement strategy is proposed, integrating mixed-scale SSM, channel enhancement branches (CEB), and frequency enhancement branches (FEB) to strengthen defect representation from both spatial and frequency domains, improving low-contrast small-target detection. Experiments on NEU-DET demonstrate that the proposed model achieves 89.7% mAP@0.5, improving 4.1 percentage points over baseline DEIM, while reducing parameters by 45.5% and floating-point operations (FLOPs) by 61.1%, reaching 112.6 frames per second (FPS) on personal computer (PC) and 61.5 FPS on RK3588 under mixed-precision quantization. Generalization experiments on PCB-DET and GC10-DET further validate robustness, with mAP@0.5 reaching 97.8% and 90.6%, respectively. These results validate that the proposed framework effectively leverages parallel computing architectures to bridge HPC-powered training and real-time edge deployment for industrial 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

Mamba-HCIM: a lightweight steel plate surface defect detector via spatial-frequency dual enhancement

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Xiaojun Hu,
  • Zhuguo Zhou,
  • Yunjie Xu

摘要

Steel plate surface defect detection remains challenged by the simultaneous demands for real-time processing, multi-scale defect identification, and robustness against complex backgrounds. Addressing these demands constitutes a computationally intensive task that inherently requires high-performance computing (HPC) capabilities, including graphics processing unit (GPU)-accelerated parallel training and heterogeneous neural processing unit (NPU)-based real-time edge inference. To address these challenges, this study proposes an improved, lightweight DETR with improved matching (DEIM) framework that systematically integrates state space model (SSM)-based global modeling, wavelet-domain edge enhancement, hierarchical cross-spatial recalibration, and Fourier-domain defect-background separation. Specifically, a lightweight SSM-based backbone is introduced to achieve global receptive field coverage with linear computational complexity via recursive inference. It is combined with wavelet transform and multi-kernel depthwise convolution to enhance high-frequency edge extraction and multi-scale feature representation. A hierarchical feature fusion encoding (HFFE) structure is then designed to bridge semantic gaps between different levels and suppress background texture interference through cross-spatial weight recalibration and coordinate attention. Furthermore, a spatial-frequency dual-domain feature enhancement strategy is proposed, integrating mixed-scale SSM, channel enhancement branches (CEB), and frequency enhancement branches (FEB) to strengthen defect representation from both spatial and frequency domains, improving low-contrast small-target detection. Experiments on NEU-DET demonstrate that the proposed model achieves 89.7% mAP@0.5, improving 4.1 percentage points over baseline DEIM, while reducing parameters by 45.5% and floating-point operations (FLOPs) by 61.1%, reaching 112.6 frames per second (FPS) on personal computer (PC) and 61.5 FPS on RK3588 under mixed-precision quantization. Generalization experiments on PCB-DET and GC10-DET further validate robustness, with mAP@0.5 reaching 97.8% and 90.6%, respectively. These results validate that the proposed framework effectively leverages parallel computing architectures to bridge HPC-powered training and real-time edge deployment for industrial defect detection.