Scholarlink

<p>Efficient detection of minute defects in printed circuit board assemblies (PCBAs) remains challenging due to multi-scale variation, low contrast, and strict computational constraints in industrial edge deployment. GMO-DETR, a lightweight detection framework, addresses these challenges by balancing accuracy and efficiency relative to the baseline Real-Time Detection Transformer (RT-DETR). Its architecture comprises four principal components: the Ghost-MambaOut network (GMONet) backbone, extending MambaOut into a dual-path structure to preserve fine-grained features while minimizing computational redundancy via cheap linear operations; the Context-Aware Feature Fusion (CAFF) module, which replaces standard RepC3 blocks by integrating a multi-branch RepNCSP structure with Context Anchor Attention (CAA) to capture directional context while reducing computational overhead; the Token-Aware Interaction with Feature Integration (TAIFI) module, an improved variant of Adaptive Interaction with Feature Integration (AIFI), applying token-level statistical attention to enhance subtle defect representation while streamlining encoder computations; and a GSConv-based neck optimization, replacing standard convolutions with group shuffle convolutions to lower the number of parameters (Parameters) and floating point operations (FLOPs) without compromising feature fusion. Evaluated on the PCBA-DET dataset, GMO-DETR achieves 98.27% mean average precision at 0.5 intersection-over-union (mAP@0.5), running at 86.62 frames per second (FPS) on a single GPU with 12.05 million parameters (39.4% reduction) and 40.6 GFLOPs (28.8% reduction). These results demonstrate a superior balance between detection accuracy and efficiency, confirming GMO-DETR’s suitability for high-throughput inspection in resource-constrained industrial settings.</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

GMO-DETR: a lightweight transformer for accurate and efficient PCBA defect detection

  • 期刊论文

  • 发表日期：

    2026年4月3日

查看全文
  • Weixing Su,
  • Bofan Wu,
  • Fang Liu,
  • Yudi Hu,
  • Lekang Yan

摘要

Efficient detection of minute defects in printed circuit board assemblies (PCBAs) remains challenging due to multi-scale variation, low contrast, and strict computational constraints in industrial edge deployment. GMO-DETR, a lightweight detection framework, addresses these challenges by balancing accuracy and efficiency relative to the baseline Real-Time Detection Transformer (RT-DETR). Its architecture comprises four principal components: the Ghost-MambaOut network (GMONet) backbone, extending MambaOut into a dual-path structure to preserve fine-grained features while minimizing computational redundancy via cheap linear operations; the Context-Aware Feature Fusion (CAFF) module, which replaces standard RepC3 blocks by integrating a multi-branch RepNCSP structure with Context Anchor Attention (CAA) to capture directional context while reducing computational overhead; the Token-Aware Interaction with Feature Integration (TAIFI) module, an improved variant of Adaptive Interaction with Feature Integration (AIFI), applying token-level statistical attention to enhance subtle defect representation while streamlining encoder computations; and a GSConv-based neck optimization, replacing standard convolutions with group shuffle convolutions to lower the number of parameters (Parameters) and floating point operations (FLOPs) without compromising feature fusion. Evaluated on the PCBA-DET dataset, GMO-DETR achieves 98.27% mean average precision at 0.5 intersection-over-union (mAP@0.5), running at 86.62 frames per second (FPS) on a single GPU with 12.05 million parameters (39.4% reduction) and 40.6 GFLOPs (28.8% reduction). These results demonstrate a superior balance between detection accuracy and efficiency, confirming GMO-DETR’s suitability for high-throughput inspection in resource-constrained industrial settings.