Image Super-Resolution Screening for Reliable Defect Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI-based defect inspection models for semiconductor manufacturing face challenges in generating high-quality super-resolved images due to dependency on input quality and resolution, leading to inefficient resource usage and unreliable defect detection.
Innovation Solution
A multitask learning model is employed to determine image properties and select suitable images for super-resolution, recommending parameters for upscaling, thereby minimizing resource consumption and enhancing defect detection reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution images are acquired using imaging techniques, then defect detectability is improved, but acquisition time and resource consumption increase significantly
Solution Approach 1:
The system performs preliminary assessment of low-resolution images to identify candidate regions that warrant super-resolution processing. By pre-screening images and regions before applying computationally intensive SR algorithms, the system avoids processing all images at high resolution, thus reducing overall acquisition and processing time while maintaining defect detectability for critical regions.
Solution Approach 2:
Instead of uniformly processing all images at high resolution, the system applies super-resolution selectively to specific regions of interest or candidate defect regions identified in low-resolution images. This local approach concentrates computational resources on areas where high resolution is most beneficial for defect detection, reducing overall resource consumption while maintaining measurement precision where needed.
2Adaptability or versatility
If super-resolution models are trained under stringent conditions, then generalizability is improved, but dependency on input quality and parameters increases
Solution Approach 1:
The system dynamically adjusts super-resolution processing parameters based on the characteristics of each input image or region. Rather than using fixed stringent training conditions, the system adapts resolution targets, processing intensity, and model selection based on detected features, noise levels, and region importance, thereby maintaining reliability across varying input qualities while preserving generalizability.
Solution Approach 2:
The system changes processing parameters adaptively based on input image properties. Different upscale factors, model architectures, or processing strategies are selected depending on the input resolution, noise level, and detected feature characteristics. This parameter adaptation allows the system to maintain reliable output quality across diverse input conditions without requiring all models to be trained under uniformly stringent conditions.
3Productivity
If low-resolution images are used for inspection, then resource consumption is reduced, but defect detectability deteriorates
Solution Approach 1:
The system segments the inspection process into multiple stages: initial low-resolution screening, candidate region identification, and targeted super-resolution processing. This segmentation allows most of the inspection workflow to operate efficiently on low-resolution images, preserving resource efficiency, while applying high-resolution processing only where necessary to maintain defect detectability.
Solution Approach 2:
The system applies super-resolution processing partially rather than universally. By identifying and processing only those regions or images where defects are suspected or where high resolution is critical for inspection decisions, the system achieves adequate defect detectability without the excessive resource consumption of processing all images at high resolution.
Data Source
AI summary
A method and system for screening a plurality of images for performing super-resolution (SR) are provided. The method includes using a multitask learning model (MLM) determining at least one of a plurality of image properties related to each of a plurality of images having a resolution lower than a predefined threshold; and selecting, based on the at least one of the plurality of image properties, a first set of images among the plurality of images to each be respectively suitable for upscaling; recommending, using a recommendation model, at least one parameter based on the at least one of the plurality of image properties; and generating at least one super-resolution image by respectively performing a super-resolution upscaling operation on at least one image of the first set of images based on the at least one parameter and the at least one of the plurality of image properties.