Title: Enhancing PCB Defect Detection via Structure-Guided Mixed Masked Pretraining and Spatial Continuity Regularization

Abstract:

Automated optical inspection (AOI) relies heavily on printed circuit board (PCB) defect detection, a task that presents significant practical hurdles due to the prevalence of minute, low-contrast anomalies obscured by dense circuitry. To overcome these challenges, this study introduces a two-stage framework for PCB defect detection that integrates spatial continuity regularization with structure-guided mixed masked pretraining. During the pretraining phase, we employ a sparse convolutional masked pretraining strategy to leverage unlabeled PCB imagery. This approach utilizes structure-guided mixed masking to generate informative masked inputs. By suppressing invalid responses within masked areas, the sparse convolutional reconstruction pipeline allows the detector’s backbone to reconstruct missing structural elements based on visible conductive patterns, effectively capturing PCB structural priors. In the subsequent fine-tuning phase, the pretrained backbone is adapted for the specific defect detection task. We introduce a spatial continuity regularization term to refine predictions by consolidating dispersed positive outputs associated with a single defect instance, thereby encouraging tighter localization along elongated defect areas. Experimental results on the DsPCBSD+ dataset demonstrate that our method reaches an mAP0.5 of 85.5% and an mAP0.5:0.95 of 52.3%, surpassing multiple robust baseline detectors. Further ablation studies and qualitative analyses validate the framework’s efficacy in ensuring robust defect detection within industrial AOI environments.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC