A SelectiveNet-Based Method for Defect Classification in Semiconductor Manufacturing

In semiconductor manufacturing, accurately classifying defects in scanning electron microscope (SEM) images is crucial for optimizing the production process. This paper introduces a novel defect classification frame-work based on SelectiveNet, which can reject predictions for defects with a high risk of misclassification through selective learning, effectively balancing the trade-offs between prediction coverage and accuracy for highly diverse and complex real SEM images. The efficacy of our proposed approach is demonstrated in the experiments with 95.14% classification accuracy with no reject and 98.37% at a selective learning coverage of 91.17% (rejection of 8.83%).