Image-Intrinsic Priors for Integrated Circuit Defect Detection and Novel Class Discovery via Self-Supervised Learning

TL;DR

The paper tackles IC SEM defect detection and novel class discovery under open-world, data-scarce conditions. It introduces IC-DefectNCD, a support-set-free framework that leverages single-image intrinsic priors through an NI-Extractor and NIG-Decoder to localize defects via reconstruction residuals, followed by self-saliency adaptive binarization for defect-centered crops. For classification, SMG-ViT incorporates a soft-mask spatial prior into a teacher–student self-supervised system and uses semi-supervised k-means to estimate the number of unseen defect classes, enabling robust recognition of unseen defects. Experiments on 2,990 SEM images across BEOL, DEP, and DPR with 15 defect types demonstrate strong detection and unseen-class discovery, and industrial deployment shows practical real-time applicability on production lines.

Abstract

Integrated circuit manufacturing is highly complex, comprising hundreds of process steps. Defects can arise at any stage, causing yield loss and ultimately degrading product reliability. Supervised methods require extensive human annotation and struggle with emergent categories and rare, data scarce defects. Clustering-based unsupervised methods often exhibit unstable performance due to missing priors. We propose IC DefectNCD, a support set free framework that leverages Image Intrinsic Priors in IC SEM images for defect detection and novel class discovery. We first develop Self Normal Information Guided IC Defect Detection, aggregating representative normal features via a learnable normal information extractor and using reconstruction residuals to coarsely localize defect regions. To handle saliency variations across defects, we introduce an adaptive binarization strategy that produces stable subimages focused on core defective areas. Finally, we design Self Defect Information Guided IC Defect Classification, which incorporates a soft mask guided attention mechanism to inject spatial defect priors into the teacher student model. This enhances sensitivity to defective regions, suppresses background interference, and enables recognition and classification of unseen defects. We validate the approach on a real world dataset spanning three key fabrication stages and covering 15 defect types. Experiments demonstrate robust performance on both defect detection and unseen defect classification.

Figures (12)

• Figure 1: Wafer site sampling and SEM imaging pipeline, with layout changes in IC manufacturing continuously introducing unseen defects.
• Figure 2: Comparison of unsupervised and supervised defect detection. (a) Unsupervised methods use normal features to localize anomalies but cannot categorize unseen defects. (b) Supervised methods classify known defects from labeled data but fail on unseen categories and demand extensive annotation.
• Figure 3: (a) Self-Normal Information Guided IC Defect Detection: NI-Extractor gathers normal informations and NIG-Decoder reconstructs patch tokens by these normal informations, and feature comparison yields a defect score map.(b) Self-Saliency-driven Adaptive Binarization: adaptive thresholding and center cropping produce a compact soft mask.(c) Self-Defect Information Guided IC Defect Classification: SMG-ViT uses the soft mask within teacher–student self-supervised learning, and semi-supervised k-means estimates the number of unknown classes and sets the number of classification heads for unseen defect classification.
• Figure 4: (a) Overview of the proposed Self-Normal Information Guided IC Defect Detection framework. The framework enhances detection robustness by mining normal cues directly from a single image without relying on external support sets. It reconstructs normal patterns guided by these cues and isolates abnormal regions through residual analysis. (b) Detailed architecture of each layer in the NI Extractor. (c) Detailed architecture of each layer in the NIG Decoder.
• Figure 5: Binarization results under different thresholds.