Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DECOR: Deep Embedding Clustering with Orientation Robustness

Fiona Victoria Stanley Jothiraj, Arunaggiri Pandian Karunanidhi, Seth A. Eichmeyer

TL;DR

DECOR tackles unsupervised wafer-map clustering under orientation variability and multi-defect patterns by integrating a rotation- and flip-invariant embedding (RCAE), a non-parametric clustering method (DeepDPM), and ensemble outlier detection. The RCAE uses dihedral symmetry $D_4$ and GroupPooling to yield a compact $128$-dimensional latent representation, while DeepDPM infers the number of clusters adaptively. On the MixedWM38 dataset, DECOR achieves superior clustering quality (NMI and ARI) compared with parametric baselines, and demonstrates orientation-robust clustering of rotated defect patterns. This approach offers a scalable, annotation-free solution for automated visual inspection in semiconductor manufacturing and potentially other domains.

Abstract

In semiconductor manufacturing, early detection of wafer defects is critical for product yield optimization. However, raw wafer data from wafer quality tests are often complex, unlabeled, imbalanced and can contain multiple defects on a single wafer, making it crucial to design clustering methods that remain reliable under such imperfect data conditions. We introduce DECOR, a deep clustering with orientation robustness framework that groups complex defect patterns from wafer maps into consistent clusters. We evaluate our method on the open source MixedWM38 dataset, demonstrating its ability to discover clusters without manual tuning. DECOR explicitly accounts for orientation variations in wafer maps, ensuring that spatially similar defects are consistently clustered regardless of its rotation or alignment. Experiments indicate that our method outperforms existing clustering baseline methods, thus providing a reliable and scalable solution in automated visual inspection systems.

DECOR: Deep Embedding Clustering with Orientation Robustness

TL;DR

DECOR tackles unsupervised wafer-map clustering under orientation variability and multi-defect patterns by integrating a rotation- and flip-invariant embedding (RCAE), a non-parametric clustering method (DeepDPM), and ensemble outlier detection. The RCAE uses dihedral symmetry and GroupPooling to yield a compact -dimensional latent representation, while DeepDPM infers the number of clusters adaptively. On the MixedWM38 dataset, DECOR achieves superior clustering quality (NMI and ARI) compared with parametric baselines, and demonstrates orientation-robust clustering of rotated defect patterns. This approach offers a scalable, annotation-free solution for automated visual inspection in semiconductor manufacturing and potentially other domains.

Abstract

In semiconductor manufacturing, early detection of wafer defects is critical for product yield optimization. However, raw wafer data from wafer quality tests are often complex, unlabeled, imbalanced and can contain multiple defects on a single wafer, making it crucial to design clustering methods that remain reliable under such imperfect data conditions. We introduce DECOR, a deep clustering with orientation robustness framework that groups complex defect patterns from wafer maps into consistent clusters. We evaluate our method on the open source MixedWM38 dataset, demonstrating its ability to discover clusters without manual tuning. DECOR explicitly accounts for orientation variations in wafer maps, ensuring that spatially similar defects are consistently clustered regardless of its rotation or alignment. Experiments indicate that our method outperforms existing clustering baseline methods, thus providing a reliable and scalable solution in automated visual inspection systems.

Paper Structure

This paper contains 16 sections, 2 equations, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Method
  3. Feature Extractor
  4. Non-Parametric Clustering
  5. Outlier Detection
  6. Experiment
  7. Dataset
  8. Results
  9. Conclusion
  10. Experimental Setup
  11. Outlier Detection
  12. Isolation Forest (IF) and Local Outlier Factor (LOF)
  13. Robust Cut and Thresholding
  14. Adaptive Thresholding and LOF Parameters
  15. Ensemble Outlier Detection
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Overview of the workflow. Input wafer images are processed through an embedding extractor model, generating 128-dimensional embeddings along with their reconstructed images. These embeddings are then fed into a trained non-parametric clustering model based on DeepDPM Ronen:CVPR:2022:DeepDPM to obtain cluster assignments. Finally, embeddings within each cluster are analyzed using an ensemble outlier detection algorithm to flag potential outliers.
  • Figure 2: Examples of MixedWM38 images clustered together by DECOR. Each panel represents a different cluster. The grouping of images with varying rotational orientations (scratch, and local pattern) indicates that DECOR exhibits rotational invariance.
  • Figure 3: Examples of clusters produced by DECOR with detected outliers highlighted in red. (a) a center defect cluster containing instances of donut defect outliers (b) a local defect cluster containing instances of random defect outliers (c) a donut defect cluster containing instances of random and nearfull defect outliers (d) a donut + scratch defect cluster containing instances of donut defect outliers.