Hyperspectral Detection Device Using Neural Network

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing hyperspectral detection methods require complex and costly algorithms for reconstructing three-dimensional images from two-dimensional compressed images, which do not contain additional spatial or spectral information, and are not suitable for direct detection of particularities in non-homogeneous scenes.

Innovation Solution

A device using a network of deep and convolutive neurons is employed to directly detect specificities in a compressed, non-homogeneous, two-dimensional representation of a three-dimensional hyperspectral scene, eliminating the need for image reconstruction and enabling real-time detection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If complex reconstruction algorithms are used to obtain three-dimensional hyperspectral images from two-dimensional compressed images, then measurement precision is improved, but device complexity and computational cost increase significantly

Engineering Contradiction:
Improvedetection accuracyVSAvoidalgorithm complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The invention extracts only the necessary detection information directly from the compressed two-dimensional image using a neural network, rather than reconstructing the complete three-dimensional hyperspectral image. This extraction approach obtains the required spectral and spatial features without the computational burden of full reconstruction algorithms.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The invention replaces complex mathematical reconstruction algorithms with a neural network-based detection system. The neural network is trained to directly detect features in compressed images, substituting iterative matrix inversion and reconstruction processes with a learned detection model that operates efficiently on compressed data.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Loss of information

If three-dimensional hyperspectral image reconstruction is performed, then spatial and spectral information is obtained, but processing time increases and real-time detection is not achieved

Engineering Contradiction:
Improvespatial and spectral informationVSAvoidprocessing time
Core Design Contradiction:
Loss of informationVSLoss of time

Solution Approach 1:

The neural network is pre-trained on a large dataset of compressed hyperspectral images and their corresponding labels. This preliminary training enables the network to directly detect features in new compressed images without requiring time-consuming reconstruction processes, achieving real-time detection while preserving necessary spatial and spectral information.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The invention processes data in the compressed two-dimensional domain rather than reconstructing to three-dimensional space. The neural network operates directly on the compressed image representation, detecting features in this alternative dimensionality without needing to transform back to the original three-dimensional hyperspectral volume.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Measurement precision

If compressed images are stored for later processing, then detection accuracy is maintained, but memory requirements and storage costs increase

Engineering Contradiction:
Improvedetection accuracyVSAvoidstorage requirements
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The neural network is designed to perform detection directly on compressed images with the same efficiency and accuracy as it would on reconstructed images. This self-sufficiency eliminates the need to store large amounts of reconstructed image data, as the detection system can operate effectively on the compact compressed representation.

Inventive Principle:
Principle #25Self-service

4Productivity

If standard deep convolutional neural networks are used for processing, then processing speed is improved, but they cannot handle non-homogeneous compressed images with non-linear particularities

Engineering Contradiction:
Improveprocessing speedVSAvoidcompatibility with non-homogeneous data
Core Design Contradiction:
ProductivityVSAdaptability or versatility

Solution Approach 1:

The neural network architecture incorporates specialized layers and processing mechanisms that adapt to local variations in the compressed image data. Different parts of the network handle different types of information (spatial vs. spectral) with appropriate processing strategies, enabling the system to manage non-homogeneous data structures while maintaining high processing speed.

Inventive Principle:
Principle #3Local quality

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

Enables real-time detection of scene particularities without storing compressed images, reducing computational complexity and cost, and providing direct detection of spatial and spectral information in non-homogeneous scenes.

Implementation Method 1

capture a diffracted image of the focal plane of the observed scene by means of a diffraction grating placed upstream of a digital sensor

Methodology Applied
Scientific EffectDiffraction: Diffraction

Data Source

PatentEP3714399B1Hyperspectral detection device
Publication Date: 2025.04.30 CARBON BEE
  • EP3714399B1 patent drawingFigure 1~2
  • EP3714399B1 patent drawingFigure 3~4
  • EP3714399B1 patent drawingFigure 5~6

AI summary

The invention relates to a device for detecting features in a three-dimensional hyperspectral scene (3), comprising a system for direct detection (1) of features in the hyperspectral scene (3) which incorporates a deep and convolutional neural network (12, 14) designed to detect the one or more searched features in the hyperspectral scene (3) from a compressed image of said hyperspectral scene.