Super-Field-of-View for Defect Inspection on Large-Sized Transparent Materials With Bernoulli Compressed Sensing and Neural Networks

Large-sized transparent materials are key structural members of many devices, and any tiny defects are serious safety hazards. Automatic optical inspection is the popular method for defect inspection, but it is hard to balance between a large field-of-view and a reasonable imaging resolution. This paper proposes using compressed sensing (CS) and neural networks to achieve super-field-of-view (sFOV) defect inspection. By designing special Bernoulli measurement matrices with CS, this method can achieve ten times of FOV with the same amount of data as traditional imaging methods. By introducing measurement matrix embedding and a neural network that supports multiple measurement rates in a single training, this method achieves defect inspection that works much better under different FOVs than state-of-the-art methods. A structured light defect detection system for transparent polyvinyl chloride specimens with three kinds of defects is built, which obtains a dataset with high-contrast images for tiny defects. The test results show spreading the sampling pixels to ten times of region size still achieves more than 90% of detection accuracy, which is 20% higher than the state-of-the-art methods under different FOVs. The performance gain becomes even more significant with the increase of FOV. So, this method can be applied to enlarge ten times of FOV with the same amount of data for scanning sensor-based detection tasks in both industrial and medical applications.