SEM images of semiconductor wafers contain defects with very different appearances and frequencies. Some categories have few labeled examples; others vary greatly in size and shape. A conventional classifier may also assign an unfamiliar anomaly to a known class even when the match is poor.
Instead of treating classification as one prediction, we implemented separate stages for locating unusual regions → assessing their characteristics → classifying known defects → flagging unfamiliar patterns. Each stage addresses a different source of uncertainty, allowing the analysis to be tailored to the visual characteristics of the target defect.
Classifying Defects with Sparse and Uneven Training Data
- Class frequencies are highly imbalanced
Rare defects may have too few labeled examples for a standard classification approach to perform reliably. - Appearance varies within a defect category
Size, shape and imaging differences can make examples from the same class look quite different. - Unknown anomalies challenge closed-set classifiers
A model trained only on predefined labels may still force an unfamiliar anomaly into one of those labels instead of identifying uncertainty.
A Multi-Stage, Defect-Aware Analysis Method
- Locate candidate anomalous regions
We identified image areas that differ from the expected normal appearance before attempting to name the defect type. - Analyze geometry and regional characteristics
We assessed the size, shape and other properties of extracted regions to support an initial distinction between broad anomaly types. - Route classification by defect characteristics
We used methods suited to each defect type, such as shape analysis for scratches and visual features for foreign particles. - Flag low-fit anomalies as potentially unknown
We treated regions that do not sufficiently match defined categories as candidates for further inspection instead of automatically assigning a known defect label.
Key Engineering Decisions
- Decompose the task before selecting models
Localization, region characterization and class assignment solve different problems and can be optimized separately. - Design around limited labeled examples
Anomaly-first processing makes greater use of the distinction from normal appearance when examples of every defect class are scarce. - Preserve an unknown category
A system should indicate when the known classes are insufficient rather than imply certainty for every abnormal region.
Where a Multi-Stage Approach Is Useful
This method is relevant to SEM-based inspection of semiconductor wafers and electronic components where defect types are unevenly represented and previously unseen anomalies need to be flagged for expert review. Its accuracy and operating thresholds must be evaluated against the target images and inspection criteria.
Matching the Analysis Method to the Defect
We implemented anomaly localization, region-specific analysis and an explicit path for potentially unknown defects. Separating these tasks allowed the processing method to reflect the defect’s visual characteristics and the limits of the available training data.