Hyperspectral Image Dimensionality Reduction via Pixel Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Hyperspectral and multispectral image data processing is challenging due to the curse of dimensionality, making clustering difficult and visualization on RGB screens impractical, as most methods rely on meaningless distance measures in high-dimensional datasets.
Innovation Solution
A computer-implemented method determines secondary values for each pixel location based on similarity measures, using coding vectors and techniques like maximum likelihood estimation and self-organizing maps, to reduce dimensionality and preserve structural relationships between pixels, facilitating clustering and visualization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If hyperspectral or multispectral image data is processed using traditional dimensionality reduction methods, then the data can be visualized on RGB screens, but the structural relationships and similarity between pixels are lost
Solution Approach 1:
The patent transforms the high-dimensional spectral data into a low-dimensional embedding space while preserving pairwise distances between pixels. This is achieved by learning a mapping function that projects pixels from hyperspectral space (hundreds of wavelengths) to a 2D or 3D embedding space where structural relationships are maintained, enabling both visualization and preservation of similarity information
Solution Approach 2:
The patent introduces an intermediate embedding space that acts as a mediator between the original high-dimensional spectral data and the final RGB visualization. This embedding space preserves the geometric structure and similarity relationships while being compatible with standard display devices, thus serving as a bridge that maintains information integrity throughout the transformation process
2Measurement precision
If clustering algorithms are applied to high-dimensional hyperspectral data, then spectral prototypes can be identified, but the curse of dimensionality makes clustering difficult and inaccurate
Solution Approach 1:
The patent performs dimensionality reduction and structure-preserving embedding before applying clustering algorithms. By pre-processing the data to transform it into a lower-dimensional embedding space that preserves pairwise distances, the patent eliminates the curse of dimensionality issue and enables clustering algorithms to work efficiently and accurately on the transformed data
Solution Approach 2:
The patent changes the parameter space by transforming data from hundreds of spectral wavelength dimensions to a compact embedding space with 2-3 dimensions. This parameter transformation maintains the essential similarity relationships while making the data suitable for standard clustering algorithms, thereby improving both accuracy and computational efficiency
3Loss of information
If high-dimensional spectral data is used for each pixel, then comprehensive compositional information is captured, but processing and analysis become computationally intractable
Solution Approach 1:
The patent extracts the essential structural relationships and similarity information from the high-dimensional spectral data and represents them in a compact embedding space. By extracting only the necessary geometric and similarity information rather than processing all original spectral dimensions, the patent achieves fast processing while preserving the essential compositional information needed for analysis
Data Source
AI summary
The disclosure concerns processing hyperspectral or multispectral images. Image data comprises a sampled image spectrum represented by first values for each pixel location representative of an intensity associated with a wavelength index. A processor determines for each pixel location second values based on a measure of similarity between pixel locations with respect to the first values such that two pixel locations that 5 are similar with respect to the first values are also similar with respect to the second values. The processor then stores for each pixel location the determined one or more second values associated with that pixel location on a data store. This way, the image data is made suitable for applications, such as clustering or displaying, while pixels that are similar in the input image are also similar in the output data. This means that a 10 structure between the pixels in the input image is preserved.