SABi-YOLO: static axial-bias modelling improves the localization of elongated PCB defects

Defects such as open circuits, short circuits and spurs in printed circuit boards (PCBs) are typically small and often extend horizontally or vertically along routing directions, making them difficult for conventional convolutional detectors to localize consistently at stringent intersection-over-union thresholds. Here, we propose SABi-YOLO (Static Axial-Bias YOLO), which introduces a directional inductive bias aligned with PCB routing structures through static dual-axis convolution (SDA). SDA combines parallel local 3 × 3, horizontal 1 × 3 and vertical 3 × 1 convolutional branches, followed by a 1 × 1 convolution that fuses local texture information with bidirectional axial features. Its lightweight implementation further employs depthwise-separable directional convolutions to reduce the parameter count and theoretical computational cost. SDA was integrated into multiple C2f stages of the YOLOv8m backbone and evaluated using a fixed DeepPCB data split, five paired random seeds, validation-based model selection and a locked test set. SABi-YOLO achieved an mAP50–95 of 0.7516 ± 0.0216, with a paired mean improvement of 0.0613 over YOLOv8m (95% confidence interval, 0.0075–0.1150), while reducing the parameter count and GFLOPs by 5.42% and 6.44%, respectively. Under the same evaluation protocol, the DSConv control did not show a stable additional benefit, consistent with the interpretation that static axial inductive bias improves localization at stringent IoU thresholds. SABi-YOLO therefore provides a favourable trade-off between localization quality and theoretical complexity on DeepPCB, although its cross-dataset applicability and efficiency on real-world edge hardware require further validation.