<p>Efficient and accurate defect detection on printed circuit boards (PCBs) is critical to product quality assurance. In this paper, we propose an optimized lightweight target detection model for PCB defect detection, aiming to solve the dual challenges of the difficulty of detecting tiny defects and complex background interference, while meeting the industrial real-time detection requirements. The method uses an uncertainty-aware adaptive training sample selection strategy (UATSS) to improve the training efficiency and detection performance of the model; introduces a detail-enhanced convolution (DEConv) module to improve the feature extraction capability of small defects; proposes a shared lightweight detail-enhanced detection head (SLDECD) to reduce the computational complexity of the model; and uses an improved loss function to improve the training stability. The experimental results show that the model in this paper achieves 97.8% mAP and 99.6% accuracy on the PCB defect dataset, which are 4.0 and 1.9 percentage points higher than the benchmark model, respectively, and at the same time reduces the model size to only 3.8 M, the computation volume from 6.2G FLOPs to 2.7G FLOPs, which is 56.5% less compared to the original model, and the number of frames per second (FPS) reached 144.1. The model has been successfully deployed on the RK3568 platform, realizing the requirements for high-precision real-time detection and embedded device deployment.</p>