A defect detection method based on improved YOLOv5s was proposed aiming at the problems that exist in printed circuit board (PCB), such as small defect targets that are not easy to be identified, high false detection rate, and large model size that is not easy to be deployed. Hierarchical density-based clustering k-means (hierarchical density-based K-means, HDBK-means) algorithm was used to regroup to get the anchor frame more suitable for the characteristics of PCB defects. The feature extraction module in the YOLOv5s backbone was replaced with a reparameterized normalized cross stage partial efficient layer aggregation network (RepNCSPELAN) improved by spatial and channel reconstruction convolution (SCConv), which greatly improved the speed of model inference while ensuring accuracy. Then the recognition ability of the model for defective targets at various scales can be improved by introducing reparameterized detail-enhanced generalized feature pyramid network (RDEGFPN) for feature fusion while reducing the computational resource consumption. Dynamic upsample (DySample) was used to innovate the feature fusion network to form generalized dynamic feature pyramid network (GDFPN), which improved the lightweight and efficiency of the model and made the model easier to deploy. Comparative experiments conducted on public PCB datasets demonstrate that the proposed algorithm achieves a 3.8% improvement in mean average precision (mAP), a 2.9% enhancement in precision, and a 26.9% reduction in model size, while achieving a detection speed of 138.1 frames per second. The deployment of the model on the RK3568 platform meets the requirements for real-time detection and embedded device implementation.