Abstract: Printed Circuit Board (PCB) defect detection is a critical link in the quality control of electronic manufacturing, and its detection accuracy and real-time performance directly determine product reliability and production efficiency. Since PCB defects are generally tiny in size and easily blended into complex circuit backgrounds, existing detection methods still have shortcomings in small object feature representation, background interference suppression and precise localization. Accordingly, this paper proposes CFSC-DETR, a detection model with cross-scale feature fusion and state-space collaborative modeling based on RT-DETR. The DualConv module is integrated into the backbone network ResNet18 to reduce computational redundancy and improve the capability of local texture extraction; in the encoding stage, a Dual Attention State-space Inference Module (DASM) is presented, which combines window attention and state-space modeling to enhance the representation of PCB topological structures and long-range contextual information and alleviate background interference; a Lightweight 3D Adaptive Cross-scale Fusion Neck Network (L3D-AFFN) is constructed in the neck network to preserve fine-grained semantics during the cross-scale fusion of tiny defects. Meanwhile, a Wasserstein-Geometric Synergistic IoU loss function (WGS-IoU) is proposed, which applies joint constraints of distribution similarity and geometric consistency to boost the bounding box regression accuracy for tiny and irregular defects. Experimental results reveal that compared with RT-DETR, the proposed model achieves improvements of 3.29, 4.06 and 4.71 percentage points in mAP50, precision and recall on the PKU-Market-PCB dataset while reducing model parameters and computational complexity, and it also exhibits favorable cross-scene generalization ability on the DsPCBSD+ and DeepPCB datasets. Compared with other mainstream object detection models, this method achieves superior overall performance for tiny PCB defect detection under complex backgrounds.