Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploring "Many in Few" and "Few in Many" Properties in Long-Tailed, Highly-Imbalanced IC Defect Classification

Hao-Chiang Shao, Chun-Hao Chang, Yu-Hsien Lin, Chia-Wen Lin, Shao-Yun Fang, Yan-Hsiu Liu

TL;DR

This work tackles real-world IC defect classification under extreme class imbalance and intra-class diversity by introducing IC-Defect-14 and ReCAME-Net. The model employs a multi-expert framework with regional-channel attention, Attraction-Repulsion-Balanced loss, metric-learning terms, hard category mining, and knowledge distillation to preserve intra-class compactness while enforcing inter-class separation, all without relying on distribution priors. Empirical results show state-of-the-art performance on IC-Defect-14 and competitive results on ImageNet-LT and iNaturalist-2018, with detailed ablations confirming the contribution of each component. The approach has practical impact for high-yield IC manufacturing, offering robust defect classification across highly skewed data distributions and complex production-line variations.

Abstract

Despite significant advancements in deep classification techniques and in-lab automatic optical inspection models for long-tailed or highly imbalanced data, applying these approaches to real-world IC defect classification tasks remains challenging. This difficulty stems from two primary factors. First, real-world conditions, such as the high yield-rate requirements in the IC industry, result in data distributions that are far more skewed than those found in general public imbalanced datasets. Consequently, classifiers designed for open imbalanced datasets often fail to perform effectively in real-world scenarios. Second, real-world samples exhibit a mix of class-specific attributes and class-agnostic, domain-related features. This complexity adds significant difficulty to the classification process, particularly for highly imbalanced datasets. To address these challenges, this paper introduces the IC-Defect-14 dataset, a large, highly imbalanced IC defect image dataset sourced from AOI systems deployed in real-world IC production lines. This dataset is characterized by its unique "intra-class clusters" property, which presents two major challenges: large intra-class diversity and high inter-class similarity. These characteristics, rarely found simultaneously in existing public datasets, significantly degrade the performance of current state-of-the-art classifiers for highly imbalanced data. To tackle this challenge, we propose ReCAME-Net, which follows a multi-expert classifier framework and integrates a regional channel attention module, metric learning losses, a hard category mining strategy, and a knowledge distillation procedure. Extensive experimental evaluations demonstrate that ReCAME-Net outperforms previous state-of-the-art models on the IC-Defect-14 dataset while maintaining comparable performance and competitiveness on general public datasets.

Exploring "Many in Few" and "Few in Many" Properties in Long-Tailed, Highly-Imbalanced IC Defect Classification

TL;DR

This work tackles real-world IC defect classification under extreme class imbalance and intra-class diversity by introducing IC-Defect-14 and ReCAME-Net. The model employs a multi-expert framework with regional-channel attention, Attraction-Repulsion-Balanced loss, metric-learning terms, hard category mining, and knowledge distillation to preserve intra-class compactness while enforcing inter-class separation, all without relying on distribution priors. Empirical results show state-of-the-art performance on IC-Defect-14 and competitive results on ImageNet-LT and iNaturalist-2018, with detailed ablations confirming the contribution of each component. The approach has practical impact for high-yield IC manufacturing, offering robust defect classification across highly skewed data distributions and complex production-line variations.

Abstract

Despite significant advancements in deep classification techniques and in-lab automatic optical inspection models for long-tailed or highly imbalanced data, applying these approaches to real-world IC defect classification tasks remains challenging. This difficulty stems from two primary factors. First, real-world conditions, such as the high yield-rate requirements in the IC industry, result in data distributions that are far more skewed than those found in general public imbalanced datasets. Consequently, classifiers designed for open imbalanced datasets often fail to perform effectively in real-world scenarios. Second, real-world samples exhibit a mix of class-specific attributes and class-agnostic, domain-related features. This complexity adds significant difficulty to the classification process, particularly for highly imbalanced datasets. To address these challenges, this paper introduces the IC-Defect-14 dataset, a large, highly imbalanced IC defect image dataset sourced from AOI systems deployed in real-world IC production lines. This dataset is characterized by its unique "intra-class clusters" property, which presents two major challenges: large intra-class diversity and high inter-class similarity. These characteristics, rarely found simultaneously in existing public datasets, significantly degrade the performance of current state-of-the-art classifiers for highly imbalanced data. To tackle this challenge, we propose ReCAME-Net, which follows a multi-expert classifier framework and integrates a regional channel attention module, metric learning losses, a hard category mining strategy, and a knowledge distillation procedure. Extensive experimental evaluations demonstrate that ReCAME-Net outperforms previous state-of-the-art models on the IC-Defect-14 dataset while maintaining comparable performance and competitiveness on general public datasets.
Paper Structure (21 sections, 10 equations, 7 figures, 9 tables)

This paper contains 21 sections, 10 equations, 7 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Industrial Classification Missions
  4. Highly-imbalanced and Long-tailed Data Classification
  5. Attention mechanism
  6. Proposed Method
  7. Overview
  8. Parallel independent learning (PIL)
  9. Attraction-Repulsion-Balanced (ARB) Loss
  10. Metric Learning Loss
  11. Hard Category Mining (HCM) Loss
  12. Knowledge Distillation
  13. Total Loss and Loss Weight Selection
  14. Experimental Results
  15. Datasets
  16. ...and 6 more sections

Figures (7)

  • Figure 1: Challenges posed by a real-world highly-imbalanced IC defect image dataset. IC defect features are mixed with IC product design features, resulting in unexplored unique data properties that complicate the defect classification task. (a) The "many-in-few" property (i.e., multi-cluster head classes) stems from significant intra-class diversity in the feature space, where a single defect category exhibits multiple distinct clusters. (b) Conversely, tail classes often contain very few samples, leading to the "few-in-many" property, where a sparse number of samples are scattered across a large feature space. Figures (c) and (d) demonstrate samples of different defect categories and product lines, highlighting that these data properties arise from different defect patterns on the same IC design and the same defect pattern on different IC designs. These unique data properties, combining diverse defect patterns and overlapping design influences, significantly complicate the classification process.
  • Figure 2: Data distributions of two imbalanced datasets: (a) iNaturalist-2018 van2018inaturalist following an ideal exponential distribution, (b) ID-Defect-14 being extremely imbalanced. The vertical-axis and horizontal-axis are number of per-class samples and class indices, respectively.
  • Figure 3: Architecture of the proposed Regional Channel Attention-based Multi-Expert Network (ReCAME-Net). ReCAME-Net comprises of two stages: Parallel Independent Learning (PIL) and Knowledge Distillation (KD).
  • Figure 4: Architecture of the proposed RC-Attn module.
  • Figure 5: Detailed architectural design of RC-Attn module.
  • ...and 2 more figures