Research on PCB Surface Small Defects Detection Method Based on Improved YOLOv11

Abstract

The automated detection of minute defects on the surfaces of the printed circuit board (PCB) is crucial for ensuring the quality and reliability of electronic products. However, the current detection methods frequently encounter significant issues with fine feature loss during downsampling, making it challenging to accurately identify small targets against complex background textures. To tackle these challenges, this article introduces the YOLO-DFA detection model, which is based on an enhanced version of YOLOv11. This model features two innovative core modules: frequency-domain decoupled space-to-depth (SPD) convolution (DF-SPD Conv) and channel refinement attention (CRA). The DF-SPD module effectively decouples high-frequency defects from low-frequency backgrounds by employing spatial lossless rearrangement and an asymmetric dual-stream architecture that utilizes channel segmentation, thereby mitigating feature loss at the fundamental architectural level. Furthermore, the CRA module purifies defective signals by eliminating redundant noise in complex environments, enhancing feature interactions across channel dimensions and employing a multidimensional feature recalibration mechanism. Experimental results demonstrate that YOLO-DFA performs exceptionally well on the DeepPCB dataset across several metrics, achieving an $\text {mAP}_{{50}}$ value of 97.5%, which surpasses the YOLOv11 baseline by 2.9%. Furthermore, maintaining a robust inference speed of 22.6 frames per second (FPS) on our evaluation platform, the proposed method demonstrates a favorable balance between detection accuracy and computational efficiency, providing strong potential for real-time PCB inspection applications.