<p>Printed Circuit Board (PCB) defect detection in industrial scenarios is of great significance to many fields, such as computers and aerospace. In recent years, the YOLO series of algorithms have been widely used in target detection and have achieved great advantages. The detection performance has also been continuously improved with the iteration of versions. However, the YOLO series of algorithms rely on the Non-Maximum Suppression (NMS) processing of the backend, which increases the time and computational cost. In this paper, we propose a PCB defect detection algorithm based on a backend-free processing method. The defect detection method based on backend-free processing avoids the limitation of the anchor box of the traditional YOLO series of algorithms on the detection ability of the model, reducing the time and computational cost. In addition, the proposed attention-based scale fusion network effectively improves the detection performance of the model and enhances the feature extraction ability of small targets. In addition, we superimpose a shallow small-target detection head and introduce an attention mechanism in the detection model to improve the model's solution space ability. We evaluate the proposed detection algorithm on the public Peking University PCB dataset. The results show that our algorithm has significant advantages over other SOTA algorithms in detection accuracy and efficiency. The detection algorithm proposed in this paper achieved an accuracy of 99.7% on Peking University's dataset, representing a 2.3% improvement over YOLO12. Notably, for the "Open_circuit" and "Spurious_copper" categories, the AP values reached 99.9% and 98.9%, respectively, Marking significant increases of 4.1% and 1.2% compared to YOLO12. Furthermore, the algorithm demonstrates exceptional efficiency, with a detection time of merely 0.01195 s per image.</p>