In this paper, we present a system for the detection and classification of defects in semiconductor units. The proposed system consists of three stages: proposal generation stage, defect detection stage and refinement stage. In the proposal generation stage, changes on the target unit are detected using a novel change detection approach. In the second stage, a deep neural network is used to classify detected regions into either defective or non-defective regions. Non-defective regions are regions exhibiting allowable changes due to factors such as lighting conditions and subtle differences in manufacturing. The defect detection stage achieves up to 94.3% accuracy. In practice, defects that are smaller than a specified tolerance size are ignored by manufacturers. The tolerance size depends on the defect types and is determined based on risk factors. In order to ignore such defects, our approach includes a final refinement stage wherein the detected defects are categorized by a stacking-based ensemble classifier into different classes. The refined system achieves up to 97.88% overall detection accuracy. The presented system is immediately applicable to different types of defects on die, epoxy and substrate. Inputs of the system can be either color or grayscale images.

中文翻译：

---

基于深度学习的局部自适应统计背景建模，用于半导体单元中的缺陷检测

在本文中，我们提出了一种用于检测和分类半导体单元缺陷的系统。所提出的系统由三个阶段组成：提议生成阶段、缺陷检测阶段和细化阶段。在提议生成阶段，使用新的变化检测方法检测目标单元的变化。在第二阶段，使用深度神经网络将检测到的区域分类为有缺陷或无缺陷的区域。非缺陷区域是由于照明条件和制造中的细微差异等因素而表现出允许变化的区域。缺陷检测阶段的准确率高达 94.3%。实际上，制造商会忽略小于指定公差大小的缺陷。容差大小取决于缺陷类型并根据风险因素确定。为了忽略此类缺陷，我们的方法包括最后的细化阶段，其中检测到的缺陷由基于堆叠的集成分类器分类为不同的类别。改进后的系统可实现高达 97.88% 的整体检测精度。所提出的系统可立即适用于芯片、环氧树脂和基板上的不同类型的缺陷。系统的输入可以是彩色或灰度图像。