Patch Similarity SNR Metric for Reliable Defect Annotation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for defect annotation, deep learning model tunability, and repeatability in semiconductor inspection are inadequate, leading to unreliable defect detection and classification, particularly for subtle defects and novel defects.
Innovation Solution
A patch similarity-based signal-to-noise ratio (SNR) metric is calculated to guide annotation, improve model tunability, and enhance repeatability by correlating with defect strength, using background noise statistics and GPU acceleration for computation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual defect annotation is used, then annotation flexibility is maintained, but annotation reliability and consistency deteriorate
Solution Approach 1:
The patent introduces an SNR metric as an intermediary tool between the inspection system and annotators. This metric objectively quantifies defect strength and serves as a mediator that guides annotators in selecting appropriate defects for annotation, thereby improving reliability without significantly increasing process complexity
Solution Approach 2:
The SNR metric provides feedback to annotators about the relative strength of detected defects, enabling them to make more informed decisions about which defects to annotate. This feedback mechanism improves annotation consistency and reliability while maintaining operational simplicity
2Reliability
If deep learning models are trained with current methods, then model capability is achieved, but model robustness and repeatability deteriorate
Solution Approach 1:
The patent changes the parameter used for defect selection from subjective visual assessment to objective SNR metric values. By using SNR as the selection criterion, the patent improves model robustness and repeatability while maintaining training efficiency through automated, objective defect selection
Solution Approach 2:
The patent replaces the mechanical/subjective process of manual defect selection with an automated computational approach using SNR metrics. This substitution improves model robustness by eliminating human subjectivity while maintaining or improving training efficiency through automation
3Measurement precision
If inspection sensitivity is increased to detect subtle defects, then defect detection capability improves, but noise and false positives increase
Solution Approach 1:
The patent uses SNR metric as a parameter that quantifies the relationship between defect signal strength and background noise. By selecting defects based on SNR threshold, the patent achieves high detection precision for subtle defects while filtering out false positives through objective signal-to-noise evaluation
Data Source
AI summary
Methods and systems for determining a signal-to-noise metric for locations of interest on a specimen are provided. One or more statistics of non-defect signals from background patch images in a test image that are similar to a patch image of a location of interest in the test image are determined. The background patch images are found by searching a reference image for patch images that are similar to the location of interest patch image and finding the corresponding patch images in the test image. The signal of the location of interest in the test image and the one or more statistics are used to determine a signal-to-noise metric for the location of interest. The signal-to-noise metric can be used in applications such as defect annotation, deep learning (DL) model tunability, DL model repeatability, and novel defect detection.