Dimension Measurement Using Human Pose Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing dimension measurement methods for semiconductor devices, particularly in advanced processes, face challenges such as manual intervention, operator dependence, and increased complexity due to the need for nanometer-level accuracy and the introduction of new materials and structures.
Innovation Solution
The application of human pose estimation (HPE) techniques, which utilize machine learning models to estimate the pose of objects in images, is used to automatically extract coordinate information for dimension measurement. This involves preprocessing the training data to allow for partial annotation and integrating images with insufficient measurement portions, enabling the model to learn and adapt without requiring full annotation of all samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual dimension measurement is used, then measurement precision can be maintained, but operator dependence increases and productivity decreases
Solution Approach 1:
The patent replaces manual mechanical measurement with an automated image recognition system using deep learning. The neural network automatically detects patterns, extracts dimension information, and performs measurements without human intervention, thereby eliminating operator dependence while maintaining precision and increasing throughput.
Solution Approach 2:
The measurement system performs self-service by automatically processing images and extracting dimensional data without requiring operator intervention. The deep learning model independently completes the entire measurement workflow from image input to dimension extraction, enabling autonomous operation.
2Measurement precision
If manual dimension measurement is used, then measurement precision can be maintained, but operator dependence increases
Solution Approach 1:
The patent replaces manual mechanical measurement with an automated image recognition system using deep learning. The neural network automatically detects patterns, extracts dimension information, and performs measurements without human intervention, thereby eliminating operator dependence while maintaining precision.
3Measurement precision
If full annotation of all training samples is required, then model accuracy improves, but time consumption increases
Solution Approach 1:
The patent applies partial annotation by requiring annotation of only a subset of training samples rather than all samples. The deep learning model learns from this partial annotated data and can still achieve high accuracy, significantly reducing the time required for data preparation while maintaining model performance.
4Measurement precision
If measurement points increase for complex structures, then measurement precision improves, but device complexity increases
Solution Approach 1:
The patent replaces complex manual measurement procedures with an automated deep learning system. The neural network automatically identifies and measures multiple dimension points on complex structures without requiring manual intervention, thereby improving precision while reducing the operational complexity of the measurement system.
Data Source
AI summary
A computer system for extracting, from image data, coordinate information on base points for measuring a dimension of a desired portion of a pattern, and measuring the dimension by using the coordinate information, the computer system including a preprocessing unit configured to allow training by matching all of the samples by setting a base point insufficient in annotation data as an insufficient measurement portion and shielding the insufficient measurement portion on the image data for the sample in which coordinate values of only a part of the base points are described, the preprocessing unit is also configured to allow learning and is trained in advance by using training data in which the image data is set as an input and the coordinate information is set as an output, and the preprocessing unit extracts the coordinate information and the dimension for new image data input for learning.