Pose-guided YOLO-CED: A two-stage framework for robust SMD-PCB defect detection in mobile manipulators.

The stability of printed circuit boards (PCB) and their surface-mount devices (SMD) holds significant importance in industrial production applications. However, when deploying mobile manipulator to replace manual safety inspections in manufacturing environments, challenges persist due to uncertainties in PCB spatial pose, significant variations in SMD dimensions, and computational limitations. Achieving stable identification across all defect categories remains difficult. To address this, this paper proposes a two-stage method for SMD-PCB defect detection in mobile manipulator, combining target pose estimation with YOLO-CED. First, a color attention mechanism is introduced. By associating color features with spatial features, it adaptively assigns appropriate weights to different regions. High-precision planar detection enhances the robustness of point cloud-based pose estimation. Second, an adaptive gating mechanism with learnable parameters is introduced for feature extraction, handling channel weights for three-branch parallel feature acquisition fusion and lightweight spatial weights. Proposing a multi-sensory field feature acquisition mechanism and an edge sharpening feature fusion mechanism based on Laplace operator convolution. Concurrently, a detail restoration branch ensures precise boundary localization by preserving fine features of small objects. Finally, the spatial reshaping branch features are processed through an L2-normalized self-supervised multi-head attention mechanism, and the dual-branch features are fused to produce the final output. Experimental results demonstrate that the proposed method outperforms multiple lightweight models on the self-built HUSTPCB dataset (covering 10 common defect categories), achieving a mean average precision (mAP) of 81.7%. This algorithm was verified on a mobile platform integrated with a robotic arm, demonstrating the application prospects of the mobile manipulator in flexible PCB manufacturing. Meanwhile, its excellent performance on the PKU-Market-PCB dataset proved the universality and transferability of the YOLO-CED.