| Περίληψη: |
As electronic products continue to advance toward miniaturization, increasing complexity, and higher performance, PCB design has become more sophisticated, placing greater demands on materials, manufacturing processes, and fabrication precision. This rising complexity makes defect detection more challenging, and achieving high-accuracy, real-time inspection under cost constraints has become a central research problem. To address these challenges, we propose a lightweight and high-precision detection framework based on an improved YOLOv11 architecture, named YOLO-Semantic-Aware Fusion (YOLO-SAF).The proposed model introduces a Deformable Fusion Module (DFM) that combines multi-scale dilated and deformable convolutions to enhance edge detail and spatial adaptability. A Multi-Branch Aware Feature Pyramid Network (MAFPN) is designed to preserve fine-grained details in shallow layers while enriching semantic abstraction in deeper layers, enabling effective cross-scale and cross-position information flow. In addition, considering the differences between PCB defect detection and conventional scenarios, the detection head of YOLOv11 is optimized to focus on small-object regions, improving accuracy while enhancing model efficiency.Extensive experiments on the public HRIPCB dataset show that YOLO-SAF achieves 98.3% mAP with only 5.93 M parameters and 13.6 GFLOPs, demonstrating its effectiveness and deployment potential for industrial PCB defect inspection. [ABSTRACT FROM AUTHOR] |