Image Evaluation Apparatus for Semiconductor Defect Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional Neural Networks (CNN) used in semiconductor inspection often misrecognize unknown defects due to limited training data, leading to potential overlooking of fatal defects and reduced yield.
Innovation Solution
An image evaluation apparatus and method that utilizes multiple classifiers and an evaluation unit to classify defect information by comparing image-of-interest information with defect region information, preventing misrecognition by machine learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If CNN is used for defect classification in semiconductor inspection, then classification accuracy for known defects is improved, but misrecognition of unknown defects occurs frequently
Solution Approach 1:
The patent introduces an evaluation unit as an intermediary component that assesses the reliability of CNN classification results. This evaluation unit detects uncertain classifications and refers them to alternative processing methods, preventing misrecognition of unknown defects while maintaining high accuracy for known defects.
Solution Approach 2:
The system dynamically adjusts its classification approach based on the confidence level of the CNN. For high-confidence predictions, the CNN result is accepted; for low-confidence predictions, the system switches to alternative processing methods, optimizing both accuracy and reliability adaptively.
2Productivity
If training data is limited to common defect types, then CNN learning efficiency is improved, but detection of rare defects becomes difficult
Solution Approach 1:
The evaluation unit serves multiple functions: it identifies unknown defects, assesses classification confidence, and routes uncertain cases to alternative processing. This multi-functional component enables the system to handle both common and rare defects effectively without requiring extensive retraining.
Solution Approach 2:
The system performs preliminary evaluation of classification confidence before finalizing defect identification. By pre-assessing the reliability of CNN predictions, the system can prepare alternative processing methods in advance for uncertain cases, improving detection of rare defects without sacrificing learning efficiency.
3Speed
If machine learning model confidence is high, then classification speed is improved, but misrecognition of unknown defects increases
Solution Approach 1:
The evaluation unit provides feedback on the reliability of CNN classification results. This feedback mechanism allows the system to identify and correct potential misrecognitions while maintaining fast processing for reliable classifications, balancing speed and accuracy.
Solution Approach 2:
The system dynamically adjusts its processing path based on confidence assessment. High-confidence predictions are processed quickly through the CNN, while low-confidence predictions trigger additional evaluation steps, optimizing the trade-off between speed and reliability.
Data Source
AI summary
The purpose of the present invention is to provide an image evaluation device and method which can detect unknown defects and which can prevent misrecognition by a machine learning model. This image evaluation device, which uses a machine learning classifier to classify defect information in a defect image of an electronic device, is characterized by being provided with: an image storage unit which stores a defect image of an electronic device; a defect region storage unit which stores defect region information that is in the defect image; a classifier which classifies the defect information with machine learning; an image extraction unit which, in the course of the defect image classification processing, extracts image-of-interest information which the classifier will focus on; and an evaluation unit which compares the image-of-interest information and the defect region information to evaluate the classifiability of the defect image.