Quantitative analysis and segmentation method for defects of Y2O3 coating by improved U-Net

Scanning electron microscopy (SEM) is one of the most common approaches for characterizing of surface performance. Images from SEM contain detailed information about crystalline properties or defects, but they are typically analyzed through a laborious, serial process that relies on the trained eye of experts. This study explored a quantitative analysis method for evaluating defects in Y 2 O 3 steel coatings, based on an improved U-Net image segmentation model. The study evaluated several neural networks as U-Net encoders and developed two quantitative indicators based on the results from segmentation model. The proposed method was validated to rapidly and accurately quantify the defect performance of Y 2 O 3 steel coatings, achieving 85% agreement with expert assessments and saving at least one man seven days’ work to finish the analyzing task. This research demonstrates a promising application of SEM image segmentation at scale and makes a valuable attempt at using transfer learning for improving model performance. The automated analysis approach reduces human bias and enables large-scale defect quantification.