DSRD-YOLO: A Lightweight and High-Precision Framework for PCB Defect Detection

High-density printed circuit board (PCB) manufacturing requires reliable and efficient defect inspection. However, existing automated visual inspection methods are often challenged by small defect scales, complex background textures, and imaging noise. To address these issues, this paper proposes DSRD-YOLO, a lightweight detection framework based on YOLOv8. First, a Dynamic Spatial Residual Downsampling (DSRD) module is designed to preserve fine-grained spatial information during downsampling by integrating multi-branch feature extraction with dynamic convolution. Second, a Spatial Pyramid Pooling Fast module enhanced with Lightweight Directional Spatial Attention (SPPF-LDSA) is introduced to better capture directional structural features in PCB images. Third, a SlideLoss-based optimization strategy is adopted to improve training stability under hard-sample conditions, dynamically adjusting the focus on ambiguous samples near decision boundaries. Experiments on the PKU-Market-PCB dataset show that the proposed method achieves $95.49\pm 0.18\%$ mAP@0.5 (mean ± standard deviation over five independent runs) and improves recall by 4.6% compared with the YOLOv8n baseline; paired $t$ -tests confirm that the gains are statistically significant ( $p\lt 0.01$ ). The model maintains real-time performance at 170.0 FPS with 3.5M parameters and 9.2 GFLOPs. Further evaluations on DEEP_PCB and NEU_DET datasets indicate that the method generalizes reasonably to other industrial inspection scenarios. These results suggest that the proposed method offers a competitive and computationally efficient solution for industrial PCB defect detection.