Hybrid Detection-Segmentation for Precision Wire Welding with Mask-Based Offset Generation in Smart Connector Manufacturing

Accurate conductor-to-pad alignment is essential in PCB wire welding because misalignment increases rework, lowers yield, and affects product quality. In manufacturing images, precise conductor localization is challenging because the target structures are small, the boundaries are subtle, and the conductors often resemble nearby PCB pads. This study proposes an integrated vision pipeline for precision wire welding under deployment-oriented runtime requirements. The framework combines YOLOv9-Tiny for rapid conductor localization with EfficientViT-SAM for box-prompted mask refinement, enabling accurate conductor segmentation while maintaining practical inference speed. A total least squares (TLS)-based geometric method is introduced to generate x-axis correction offsets from predicted masks for alignment support. The models were validated on data from a 4-conductor system collected under production-like conditions. For box-prompted segmentation, EfficientViT-SAM-L2 achieved an mIoU of 95.9% with a mean inference time of 2.264 s, satisfying the target requirement of high segmentation accuracy and inference time below 3 s. Compared with heavier SAM variants and a transformer-based benchmark, the selected model provided a more practical balance between mask quality and computational efficiency. These results support the feasibility of mask-based offset generation for precision wire welding.