晶圓電性測試會產生晶圓圖，用以記錄每顆晶粒之測試結果。若缺陷僅零星且隨機分布，其影響通常較為有限，但若缺陷於特定區域形成群聚、環狀或局部集中等空間分布型態，往往反映製程中存在異常。因此，晶圓缺陷辨識不僅在於判斷缺陷是否存在，更可協助工程師快速篩選可疑晶圓，以理解缺陷的位置、範圍與程度，作為品質管理及製程異常追蹤的重要依據。現有研究多採用深度學習進行晶圓圖分類，雖具有良好的辨識效能，但其決策依據較缺乏可解釋性，且效能評估亦可能受到資料切分影響。有鑑於此，本研究以公開資料集WM-811K 為研究對象，把晶圓圖視為具有空間結構的二元格點資料，提出一套特徵少、複雜度低且可解釋的缺陷辨識方法。
本研究從缺陷的整體數量、發生位置、空間同質性與局部群聚四個面向萃取七項核心特徵：缺陷覆蓋率、外圈集中程度、Moran’s I、Geary’s C，以及三項SaTScan 群聚特徵，再代入Logistic Regression、Random Forest 與XGBoost 等分類模型，另以影像紋理特徵與卷積神經網路（Convolutional Neural Network, CNN）作為比較基準。模型評估採by-lot 資料切分搭配平衡抽樣，並進行300 次重複模擬，以F1-score 等指標及其分布評估模型的辨識效能與穩定性。
結果顯示，在瑕疵、非瑕疵的二分類中，本研究方法僅用七項空間統計特徵就能取得約0.95 的F1-score，高於影像紋理特徵基準的0.89，也在本研究採用之CNN baseline 下略高於CNN 的0.94，且呈現較佳的穩定性。另外，在去除Near-full 的七類缺陷多分類中，本研究方法的F1-score 約0.90，在本研究設定下亦高於CNN baseline 的約0.79，並保留較直接的特徵解釋性。整體而言，本研究透過少量空間統計特徵可歸納出晶圓圖中缺陷分布的重要資訊，而且在維持辨識效能的同時，提供直接的影像判讀依據。本文方法並非取代深度學習，而是提供一種兼具低維度、穩定性與可解釋性的替代或輔助方案，可作為智慧製造與半導體製程品質管理之實用工具。

Wafer electrical testing produces wafer maps that record the test result of individual dies. While sporadic and randomly distributed defects generally have limited impact, defects exhibiting clustered, ring-shaped, or locally concentrated spatial patterns often indicate underlying process abnormalities. Consequently, wafer defect recognition is not only concerned with determining the presence of defects but also with helping engineers rapidly identify suspicious wafers and understand the location, extent, and severity of defect patterns for quality control and process monitoring. Although deep learning has become the dominant approach for wafer map classification and generally achieves high recognition accuracy, its decision-making process is often difficult to interpret, and reported performance may be influenced by data partitioning strategies. To address these limitations, this study uses the publicly available WM-811K dataset and models wafer maps as binary lattice data with spatial structure, proposing a low-dimensional, computationally efficient, and interpretable defect recognition framework.
Seven spatial statistical features are extracted from four aspects of defect patterns: overall defect quantity, spatial location, spatial homogeneity, and local clustering. These features comprise defect coverage, outer-ring concentration, Moran’s I, Geary’s C, and three SaTScan-based clustering measures , which are used to construct Logistic Regression, Random Forest, and XGBoost classifiers. Model performance is evaluated using by-lot data partitioning, balanced sampling and 300 repeated simulations. Recognition accuracy and stability are assessed using the F1-score and its distribution, with image texture features and a convolutional neural network (CNN) serving as comparison baselines.
The proposed approach achieves an F1-score of approximately 0.95 for binary classification using only seven spatial statistical features, outperforming the image texture baseline (0.89) and slightly exceeding the CNN baseline (0.94) while demonstrating greater stability. For seven-class defect classification after excluding the Near-full catego-ry, it attains an F1-score of approximately 0.898, compared with approximately 0.79 for the CNN baseline, while retaining substantially greater interpretability. Overall, the proposed method effectively captures the essential characteristics of wafer defect distributions, providing a robust and interpretable alternative or complement to deep learning without compromising recognition performance.

高崇哲（2025），「強化深偽偵測：以統計方法辨識影像的圖像特徵」，國立政治大學統計學系學位論文。
陳慧霜（2023），「影像分析與深偽影片的偵測」，國立政治大學統計學系學位論文。
黃政嘉（2024），「Deepfake 與 GAN 真偽圖像統計分析」，國立政治大學統計學系學位論文。
Alqudah, R., Al-Mousa, A. A., Abu Hashyeh, Y., & Alzaibaq, O. Z. (2023). A systemic comparison between using augmented data and synthetic data as means of enhancing
wafermap defect classification. Computers in Industry, 145, 103809.
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.
Chen, T., & Guestrin, C. (2016, August). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785–794).
Cox, D. R. (1958). The regression analysis of binary sequences. Journal of the Royal Statistical Society: Series B (Methodological), 20(2), 215–232.
Geary, R. C. (1954). The contiguity ratio and statistical mapping. The Incorporated Statis-tician, 5(3), 115–145.
Jeong, I., Lee, S. Y., Park, K., Kim, I., Huh, H., & Lee, S. (2023). Wafer map failure pattern classification using geometric transformation-invariant convolutional neural network. Scientific Reports, 13(1), 8127.
Kang, H., & Kang, S. (2021). A stacking ensemble classifier with handcrafted and con-volutional features for wafer map pattern classification. Computers in Industry, 129,
103450.
Kudrov, M., Bukharov, K., Zakharov, E., Grishin, N., Bazzaev, A., Lozhkina, A., Se-menkin, V., Makhotkin, D., & Krivoshein, N. (2019). Classification of wafer maps de-
fect based on deep learning methods with small amount of data. In 2019 International Conference on Engineering and Telecommunication (EnT) (pp. 1–5). IEEE.
Kulldorff, M. (1997). A spatial scan statistic. Communications in Statistics - Theory and Methods, 26(6), 1481–1496.
Moran, P. A. P. (1950). Notes on continuous stochastic phenomena. Biometrika, 37(1/2), 17–23.
Piao, M., Jin, C. H., Lee, J. Y., & Byun, J.-Y. (2018). Decision tree ensemble-based wafer map failure pattern recognition based on Radon transform-based features. IEEE Trans-
actions on Semiconductor Manufacturing, 31(2), 250–257.
Saqlain, M., Jargalsaikhan, B., & Lee, J. Y. (2019). A voting ensemble classifier for wafer map defect patterns identification in semiconductor manufacturing. IEEE Transactions
on Semiconductor Manufacturing, 32(2), 171–182.
Shin, E., & Yoo, C. D. (2023). Efficient convolutional neural networks for semiconductor wafer bin map classification. Sensors, 23(4), 1926.
Shortridge, A. (2007). Practical limits of Moran’s autocorrelation index for raster class maps. Computers, Environment and Urban Systems, 31(3), 362–371.
Sokal, R. R., & Oden, N. L. (1978). Spatial autocorrelation in biology: 1. Methodology. Biological Journal of the Linnean Society, 10(2), 199–228.
Tsai, T. H., & Wang, C. Y. (2025). Wafer map defect classification using deep learning framework with data augmentation on imbalance datasets. EURASIP Journal on Image
and Video Processing, 2025(1), 6.
Wu, M. J., Jang, J. S. R., & Chen, J. L. (2015). Wafer map failure pattern recognition and similarity ranking for large-scale data sets. IEEE Transactions on Semiconductor
Manufacturing, 28(1), 1–12.
Xu, Q., Yu, N., & Essaf, F. (2022). Improved wafer map inspection using attention mechanism and cosine normalization. Machines, 10(2), 146.
Zheng, H., Sherazi, S. W. A., Son, S. H., & Lee, J. Y. (2021). A deep convolutional neural network-based multi-class image classification for automatic wafer map failure recog-
nition in semiconductor manufacturing. Applied Sciences, 11(20), 9769.