ABF-YOLO: A Fine-Grained PCB Defect Detection Framework Integrating Axial Attention and Bidirectional Feature Fusion

Printed circuit board (PCB) defect detection is essential for ensuring manufacturing reliability and quality. However, existing methods often struggle with background noise and exhibit limited effectiveness in identifying small-scale defects. This study introduces a novel framework, denoted as ABF-YOLO, to overcome these limitations. This framework enhances PCB defect detection by integrating axial attention and bidirectional feature fusion (BiFusion). In this architecture, the axial attention module is introduced to refine fine-level feature representations and increase precision in detecting minor defects. The BiFusion module is embedded to facilitate multi-scale information fusion, enhancing the adaptability of the model to varying object dimensions. Experiments on a benchmark PCB defect dataset demonstrate that ABF-YOLO attains 97.67% mAP@0.5, 95.78% precision, and 95.83% recall, outperforming YOLOv12 by margins of +8.16%, +0.56%, and +13.49%, respectively. ABF-YOLO exhibits outstanding detection precision and robustness for small-sized defects even under complex background conditions, highlighting its practical effectiveness for challenging PCB inspection scenarios.