Osäkerhetsskattning för djuplärningsbaserad klassificering i fotomasker

Demand for photomasks in semiconductor manufacturing is rising, and manufactures turn to deep neural networks to analyze deviations automatically. Reliable deployment require the models to express their uncertainty so that rare or ambiguous defects are not missed and human operators are not overloaded with false alarms. This thesis compares several methods for uncertainty quantification in multi-class defect classification setting. Evidential Deep Learning, Fisher-information-based Evidential Deep Learning and ZigZag are evaluated against a simpler ResNet-18 neural network as a baseline and ensemble of the baseline. Post-hoc methods are investigated such as calibration and conformal prediction. Experiments cover three image datasets, FMNIST, CIFAR-10, and a synthetic photomask dataset with deviations provided by the company Mycronic. The findings show that ensembles combined with temperature scaling match or exceed the baseline in accuracy and produce well-calibrated confidence scores. Among the post-hoc set-prediction techniques, Regularized Adaptive Prediction Sets yield the tightest sets at the required coverage level and outperform Adaptive Prediction Sets, Evidential Conformal Prediction, and Least-Ambiguous Classifiers. An ensemble that uses test-time augmentation and is calibrated with temperature scaling gives the best combination of accuracy and reliability. When an even lower error rate is required, applying Regularized Adaptive Prediction Sets to the calibrated ensemble produces prediction sets that meet the target coverage of the investigated conformal prediction methods.