A Hybrid Deep and Handcrafted Feature Learning Approach for Imbalanced Wafer Map Defect Classification

Reliable wafer map defect classification is essential for semiconductor yield enhancement, yet it is challenged by limited labeled samples, severe class imbalance, and lack of interpretability. This paper presents a data-centric framework that integrates conditional generative augmentation, dual-branch deep feature extraction, and hand-crafted geometric features to address these issues. We employ a class- and mask-conditioned StyleGAN2-ADA model to synthesize high-fidelity minority-class wafer maps, achieving a 1-NN accuracy of 0.504 on the WM-811K dataset. A global-local encoder with cross-attention fuses whole-wafer context and local defect details to produce discriminative deep features. Complementing these, we extract five interpretable hand-crafted features (radial/angular profiles, connected component count, eccentricity, and edge concentration) that encode domain knowledge. The deep and hand-crafted features are concatenated and fed into a shallow-enhanced MobileNetV2 classifier trained with Class-Balanced Focal Loss to emphasize rare classes. Comprehensive experiments on WM-811K demonstrate that the proposed framework achieves a macro F1-score of 93.80\%, significantly outperforming deep-only baselines and prior methods. Ablation studies confirm the synergy between data augmentation and feature fusion. The compact model size (64K parameters) and built-in interpretability make our approach suitable for industrial deployment.