Manifold-Learned Patch Image Reconstruction for Semiconductor Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image reconstruction methods for semiconductor manufacturing, such as blind deconvolution, Lucy-Richardson, and Wiener filters, are inadequate for accurately enhancing and restoring images of semiconductor specimens due to their reliance on prior knowledge and iterative algorithms, which are slow and ill-posed, limiting their effectiveness in defect detection and feature measurement.
Innovation Solution
A system utilizing a neural network (NN) for image reconstruction, which separates images into smaller patch images and projects them onto a learned manifold to reconstruct images with improved quality by aligning to feature vectors, leveraging manifold learning to enhance image clarity and reduce noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional iterative algorithms (blind deconvolution, Lucy-Richardson, Wiener filters) are used for image reconstruction, then prior knowledge can be incorporated, but the reconstruction process becomes slow and computationally intensive
Solution Approach 1:
The patent pre-trains neural networks with prior knowledge about image formation and degradation patterns before actual reconstruction. This preliminary training phase allows the network to learn optimal reconstruction strategies in advance, eliminating the need for slow iterative computations during real-time operation while maintaining high accuracy
Solution Approach 2:
The patent replaces traditional mechanical iterative algorithms with a neural network-based system. The neural network learns to perform image reconstruction through training data, substituting the step-by-step iterative mathematical operations with a learned mapping function that produces results much faster while achieving comparable or superior reconstruction quality
2Manufacturing precision
If iterative algorithms are used for image deblurring and restoration, then some image quality improvement can be achieved, but the problems remain ill-posed and convergence is slow
Solution Approach 1:
The patent substitutes complex iterative mathematical algorithms with a neural network that has been trained to solve the ill-posed image restoration problem. The neural network learns the inverse mapping from degraded to clean images through training, replacing the need for complex iterative optimization procedures and providing both simpler implementation and faster convergence
Solution Approach 2:
The patent changes the approach from adjusting algorithm parameters iteratively to learning optimal parameters during neural network training. The network automatically learns the best transformation parameters from training data, eliminating the need for manual parameter tuning and complex iterative adjustment while improving both speed and accuracy
Data Source
AI summary
Methods and systems for image reconstruction are provided. One system includes a computer system configured for separating an image generated for a semiconductor-related specimen into patch images smaller than the image. The system also includes a neural network configured for projecting at least one of the patch images to a manifold that includes feature vectors learned from training images whose image quality meets or exceeds predetermined criteria. The neural network also reconstructs the patch image from the feature vectors it aligns to on the manifold thereby generating a reconstructed patch image having one or more image qualities better than the input patch image.