Hyperspectral Kernel Vector Detection Engine
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional hyperspectral image processing systems are computationally intensive due to the need to calculate spectral covariance and kernel values for every pixel, which is inefficient when detecting known materials, as they retain unnecessary anomaly filter computations.
Innovation Solution
The approach computes a new 'kernel vector' for each spectral reference rather than per scene pixel, reducing calculations by creating a vector of detection filter weights for each reference spectrum and only computing coherence scores for pixels with detection filter scores above a threshold, thereby reducing the number of operations required.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If spectral covariance and kernel values are calculated for every pixel using conventional methods, then material detection accuracy is maintained, but computational complexity and processing time increase significantly
Solution Approach 1:
The patent segments the computational process into two distinct phases: (1) a preprocessing phase that computes the spectral covariance matrix once for the entire hyperspectral scene, and (2) a detection phase that uses precomputed reference spectrum kernels to evaluate each pixel. This segmentation eliminates redundant covariance calculations for each pixel while maintaining detection accuracy through the use of reference-based kernel matching.
Solution Approach 2:
The patent performs preliminary computation of the spectral covariance matrix and reference spectrum kernels before actual pixel-by-pixel material detection. By precomputing these components once for the entire scene or dataset, the system avoids repeating these expensive calculations for each pixel, thereby significantly reducing overall computational complexity while preserving detection precision.
2Reliability
If kernel calculation is performed for every scene pixel, then anomaly detection capability is maintained, but processing time increases unnecessarily for known signature detection
Solution Approach 1:
The patent extracts and separates the anomaly detection component (Reed-Xiaoli filter requiring per-pixel kernel computation) from the known signature detection component (Matched Filter and ACE using reference spectra). By taking out the anomaly detection capability as an optional separate function, the system can perform efficient known signature detection without the time penalty of computing kernels for every pixel, while retaining anomaly detection capability when needed.
Solution Approach 2:
The patent applies partial action by computing kernels only for reference spectra that are actually needed for known signature detection, rather than computing kernels for all possible pixel comparisons. This selective approach reduces processing time by performing only the necessary computations for the specific detection task at hand, avoiding excessive unnecessary calculations.
3Measurement precision
If Matched Filter and ACE scores are computed for all pixels, then comprehensive material detection is achieved, but computational effort becomes prohibitively intensive
Solution Approach 1:
The patent segments the detection process into a two-stage approach: first computing Matched Filter scores for all pixels to identify potential matches, then computing ACE scores only for pixels that exceed a threshold in the first stage. This segmentation reduces computational energy by avoiding unnecessary ACE calculations for pixels that are unlikely to contain the target material, while still achieving comprehensive detection through the systematic two-phase evaluation.
Data Source
AI summary
Provided are examples of a detecting engine for determining in which pixels in a hyperspectral scene are materials of interest or targets present. A collection of spectral references, typically five to a few hundred, is used in look a through a million or more pixels per scene to identify detections. An example of the detecting engine identifies detections by calculating a kernel vector for each spectral reference in the collection. This calculation is quicker than the conventional Matched Filter kernel calculation which computes a kernel for each scene pixel. Another example of the detecting engine selects pixels with high detection filter scores and calculates coherence scores for these pixels. This calculation is more efficient than the conventional Adaptive Cosine/Coherence Estimator calculation that calculates a score for each scene pixel, most of which do not provide a detection.