Scholarlink

<p>Detection of surface defects on metal directly impacts product quality and production stability. However, the existing defect detection methods commonly suffer from issues such as large model parameter counts and low detection efficiency. To address this, this paper proposes a lightweight detection network, LDS-Net. At shallow layers, a newly designed Dynamic Inception Mixer Block is embedded into the C3K2 structure to form CDIM, a C3K2 variant for fine-grained multiscale feature representation. At deeper layers, the VSSD Block is introduced into C3K2 to construct the State-Space Feature Modulation Module (SSFM) for global context aggregation. In the attention stage, EPGO is incorporated into the original self-attention mechanism of C2PSA to construct a new EPGO-guided channel self-attention mechanism, termed EPGO-CSA, which serves as the core of DCSA and reduces redundant attention computation through adaptive Top-<i>K</i> sparse channel interaction. Converse2D is directly adopted in the neck upsampling stage to improve high-frequency edge-detail recovery during multiscale fusion. Experiments demonstrate that LDS-Net achieves a mAP@0.5 of 79.3% on the NEU-DET dataset and 66.4% on the GC10-DET dataset while maintaining a lightweight scale (approximately 2.41&#xa0;M parameters, 6.0&#xa0;GFLOPs) and real-time inference capability (approximately 168&#xa0;FPS), striking a favorable balance between detection accuracy and speed, suggesting its potential applicability to industrial online inspection scenarios.</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

LDS-Net: a lightweight detection network for real-time steel surface-defect detection

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Chenyang Xue,
  • Jingyu Wang

摘要

Detection of surface defects on metal directly impacts product quality and production stability. However, the existing defect detection methods commonly suffer from issues such as large model parameter counts and low detection efficiency. To address this, this paper proposes a lightweight detection network, LDS-Net. At shallow layers, a newly designed Dynamic Inception Mixer Block is embedded into the C3K2 structure to form CDIM, a C3K2 variant for fine-grained multiscale feature representation. At deeper layers, the VSSD Block is introduced into C3K2 to construct the State-Space Feature Modulation Module (SSFM) for global context aggregation. In the attention stage, EPGO is incorporated into the original self-attention mechanism of C2PSA to construct a new EPGO-guided channel self-attention mechanism, termed EPGO-CSA, which serves as the core of DCSA and reduces redundant attention computation through adaptive Top-K sparse channel interaction. Converse2D is directly adopted in the neck upsampling stage to improve high-frequency edge-detail recovery during multiscale fusion. Experiments demonstrate that LDS-Net achieves a mAP@0.5 of 79.3% on the NEU-DET dataset and 66.4% on the GC10-DET dataset while maintaining a lightweight scale (approximately 2.41 M parameters, 6.0 GFLOPs) and real-time inference capability (approximately 168 FPS), striking a favorable balance between detection accuracy and speed, suggesting its potential applicability to industrial online inspection scenarios.