Outlier Substrate Inspection via Multi-Dimensional Feature Space

The patent transforms the defect detection problem from conventional 2D image space to a multi-dimensional feature space. Multiple features are extracted from image pixels and organized into feature vectors, creating an N-dimensional space where defects can be distinguished from normal variations. This dimensional transformation allows statistical methods to effectively separate defect signals from process variation noise, resolving the contradiction between detection precision and system complexity.

The patent changes the parameters used for defect detection from simple pixel intensity values to multiple extracted features including intensity, gradient, texture, and other characteristics. By transforming the detection parameters into a multi-dimensional feature space and applying statistical analysis, the system achieves higher detection precision while managing complexity through systematic feature selection and dimensionality reduction techniques.

The patent merges information from multiple pixels and multiple feature types into composite feature vectors. Instead of analyzing individual pixels in isolation, the method combines intensity, gradient, texture, and other features across multiple pixels to create comprehensive feature representations. This merging approach enhances feature distinction capability while maintaining processing efficiency through vectorized operations and statistical batch processing.

By organizing pixel information into multi-dimensional feature vectors rather than scalar pixel values, the patent creates additional dimensions for distinguishing defects from noise. This dimensional expansion provides more discriminatory power for identifying real defects while enabling efficient processing through statistical methods that operate on the structured feature space.

The patent performs preliminary actions by pre-defining the multi-dimensional feature space structure, selecting relevant features, and establishing statistical baselines from reference data before actual defect detection. This preliminary setup includes determining feature extraction parameters, creating reference distributions, and configuring dimensionality reduction methods. By preparing these elements in advance, the system achieves high detection reliability while reducing real-time computational complexity during actual inspection operations.

The patent manages computational complexity by carefully selecting and transforming parameters into a optimized feature space. Through feature selection, extraction, and potential dimensionality reduction, the system identifies the most discriminative features while reducing the total number of dimensions. This parameter transformation maintains statistical outlier detection effectiveness while making the computational process more manageable and efficient.