Pattern Matching for Repetitive Semiconductor Image Shifts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing pattern matching technologies struggle with discrepancies between template and search images due to differences in capturing conditions, semiconductor pattern performance, layer shifts, and manufacturing processes, leading to failed matching and prolonged algorithm improvement times, especially in repetitive patterns.
Innovation Solution
A pattern matching device with a training unit that estimates correlation images and performs training to minimize differences between input images, using a deep learning model to calculate matching shift amounts even in repetitive patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If template matching is performed using conventional methods, then matching can be performed for simple patterns, but matching fails when there is a large discrepancy in image appearances between template and search images
Solution Approach 1:
The patent transforms the template image into the frequency domain using Fourier transform, changing the representation parameters from spatial domain to frequency domain. This parameter transformation enables the template to capture global pattern characteristics and phase relationships, making matching more robust to local variations and discrepancies in image appearance caused by different capturing conditions, semiconductor pattern variations, and manufacturing process differences.
Solution Approach 2:
The patent replaces conventional spatial domain correlation operations with frequency domain convolution operations. By using the convolution theorem, the matching process substitutes direct spatial comparison with frequency domain multiplication and inverse transformation, enabling more efficient and robust pattern recognition that handles image discrepancies better while maintaining computational feasibility.
2Productivity
If conventional template matching algorithms are used, then processing is simple, but algorithm improvement takes a long time when matching fails
Solution Approach 1:
The patent performs preliminary transformation of the template image into the frequency domain before the actual matching process. This preliminary action of computing the Fourier transform and storing the frequency domain representation allows subsequent matching operations to be performed efficiently through simple element-wise multiplication and inverse transformation, eliminating the need for repeated complex spatial domain correlations and enabling faster algorithm execution and iteration.
3Manufacturing precision
If a template is generated from design data with smoothing and rounding, then the template is close to actual patterns, but it cannot handle every discrepancy between template and search images
Solution Approach 1:
The patent applies frequency domain transformation to the template generated from design data, changing its representation to capture both the smoothed geometric characteristics and the essential pattern structure. This parameter change allows the template to maintain manufacturing precision from design data while becoming adaptable to various discrepancies through frequency domain matching that is inherently more robust to spatial variations.
Data Source
AI summary
The present disclosure proposes a pattern matching device capable of achieving matching processing that is characterized by involving a learning function even for a semiconductor pattern including a repetitive pattern, in particular. The pattern matching device pertaining to the present disclosure is provided with a learning unit for estimating a first correlation image having, as pixel values thereof, numerical values representing a correlation between a first image and a second image. The pattern matching device calculates a second correlation image having, as pixel values thereof, numerical values representing the correlation between a derivative image generated from the first image and the first image, and the learning unit performs learning so as to reduce the difference between the first correlation image and the second correlation image (refer to FIG. 1).