Designing Robust Topological Features for Wafer Map Pattern Classification

In wafer map pattern classification, extracting robust features (i.e., rotating invariant and die size invariant) is critical in ensuring consistent classification outcomes. To attain this objective, this paper explores a set of graphical features with improved robustness. The binary wafer map image is first skeletonized to extract the skeleton graph which describes the significant defective patterns as disconnected subgraphs. Subsequently, graph statistics, such as density, graph diameter, etc., are extracted to summarize the topological information of the patterns. Next, the covered subregion and die-set by each sub-graph are identified. The geometrical features of the sub- region (e.g., eccentricity, circularity, area, etc.) and the robust statistical features of the die-set (e.g., yield percentage, yield ratios, etc.) are also extracted. The effectiveness of the proposed feature extraction method is justified and demonstrated based on the WM-811K dataset from TSMC. When benchmarking the proposed method with peer methods in the literature, the proposed method not only gives improved classification accuracy but also enhanced robustness.