PCBSSD: Self-supervised symmetry-aware detector for PCB displacement and orientation inspection

Component displacement or orientation frequently impacts the electrical properties of high-density printed circuit boards (PCB), while current methods heavily depend on supervised learning. Diverging from these methods, we introduce a novel self-supervised symmetry-aware detector (PCBSSD), aiming at transitioning from supervised to unsupervised learning for detecting the displacement in symmetric devices. Specifically, a view-transform based paradigm is proposed to exploit symmetry in visual objects through consistencies across different views. Through self-supervision, PCBSSD detects device displacement from their rotational symmetry and orientations from reflective symmetry. To validate PCBSSD, we present a real-world dataset named PCBMO. Experiments show that our method, without any manual annotation, achieves performance comparable to state-of-the-art supervised methods as to accuracy and speed. To our best knowledge, PCBSSD is the first unsupervised displacement inspection approach harnessing the symmetry of devices. It offers a competitive alternative, particularly effective for symmetric objects that widely present in PCB.