PSAYOLOv9c-NWD: A Polarized Self-Attention Enhanced YOLOv9-c with Normalized Wasserstein Distance for High-Precision PCB Component Detection

This paper addresses the challenge of low accuracy in small object detection of PCB components by proposing PSAYOLOv9c-NWD, an enhanced YOLOv9-c based detection algorithm. First, we replace the standard convolutional module in RepNCSPELAN4 with a polarized self-attention (PSA) mechanism, which strengthens small object feature capture through channel-space dual-branch collaborative modeling. Subsequently, the original loss function is substituted with a normalized Wasserstein distance (NWD)-based loss function to mitigate localization sensitivity for micro-scale targets. Experimental results demonstrate that the improved PSAYOLOv9c-NWD model achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$54.7 \% \text{mAP}$</tex> and 91.446 % precision, representing improvements of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4. 9 \%}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{5 \%}$</tex> respectively over the baseline model. The effectiveness of our approach is further validated through comprehensive ablation studies and comparative experiments.