CN-121677934-B - Method, device, equipment, storage medium and program product for reconstructing spectrum data cube based on compressed physical prior

CN121677934BCN 121677934 BCN121677934 BCN 121677934BCN-121677934-B

Abstract

The invention provides a method, a device, equipment, a storage medium and a program product for reconstructing a spectrum data cube based on a compressed physical priori, which relate to the technical field of spectrum reconstruction and comprise the steps of obtaining a two-dimensional measurement image formed by encoding and modulating a target to be observed through a spectrum imaging chip; the method comprises the steps of obtaining an original transmission spectrum matrix corresponding to each pixel in a spectrum imaging chip, carrying out principal component analysis on the original transmission spectrum matrix in a spectrum dimension, extracting principal component feature vectors through a singular value decomposition method, projecting the original transmission spectrum matrix into an orthogonal subspace formed by the principal component feature vectors to obtain a compressed physical prior vector, inputting the two-dimensional measurement image and the compressed physical prior vector into a depth expansion neural network, and outputting a high-dimensional spectrum data cube of the object to be observed. The invention can obviously improve the reconstruction efficiency of the spectrum data cube.

Inventors

HUANG YIDONG
CUI KAIYU
XING ZHIYANG
ZHANG WEI
FENG XUE
LIU FANG
SUN HAO
LI YONGZHUO

Assignees

清华大学

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (9)

1. A method for reconstructing a spectrum data cube based on a compressed physical prior, comprising: Acquiring a two-dimensional measurement image formed by code modulation of a target to be observed through a spectrum imaging chip; Acquiring an original transmission spectrum matrix corresponding to each pixel in the spectrum imaging chip; Performing principal component analysis on the original transmission spectrum matrix in a spectrum dimension, and extracting principal component feature vectors through a singular value decomposition method; Projecting the original transmission spectrum matrix into an orthogonal subspace formed by the principal component feature vectors to obtain a compressed physical prior vector; inputting the two-dimensional measurement image and the compressed physical prior vector into a depth expansion neural network, and outputting a high-dimensional spectrum data cube of the target to be observed, wherein the depth expansion neural network comprises a front convolution layer, a cascade module and a tail end convolution layer; the main component analysis is carried out on the original transmission spectrum matrix in the spectrum dimension, and the main component feature vector is extracted through a singular value decomposition method, which comprises the following steps: Expanding the original transmission spectrum matrix according to the dimension of the space pixels to obtain a two-dimensional matrix, wherein each row vector of the two-dimensional matrix corresponds to the spectral transmission response data of the space pixels in the spectral imaging chip; Performing decentration pretreatment of principal component analysis on the two-dimensional matrix in a spectrum dimension to obtain a decentration matrix; Calculating a covariance matrix of a spectrum dimension according to the decentration matrix; Singular value decomposition is carried out on the covariance matrix to obtain a singular value matrix; And ordering all the eigenvectors in the singular value matrix in descending order according to singular values to obtain an eigenvector sequence, and extracting principal component eigenvectors from the eigenvector sequence, wherein the descending order ordering rearranges the eigenvectors of the spectrum dimension from high to low according to the information contribution degree of the eigenvectors.
2. The method for reconstructing a spectrum data cube based on compressed physical priors according to claim 1, wherein before inputting the two-dimensional measurement image and the compressed physical priors vector into a depth-expanded neural network and outputting the high-dimensional spectrum data cube of the object to be observed, further comprises: Constructing a training sample set of an initial depth expansion neural network, wherein the training sample set comprises a plurality of groups of training samples, and each group of training samples comprises a two-dimensional measurement sample image, a compressed physical prior sample vector and a real high-dimensional spectrum data cube; Randomly selecting a target training sample from the training sample set; Inputting the two-dimensional measurement sample image and the compressed physical prior sample vector of the target training sample into a depth expansion neural network to obtain a predicted high-dimensional spectrum data cube; Calculating a spectral distribution constraint loss value based on the predicted high-dimensional spectral data cube and the real high-dimensional spectral data cube; and training the initial deep expansion neural network based on the spectrum distribution constraint loss value to obtain a trained deep expansion neural network.
3. The method for reconstructing a spectrum data cube based on compressed physical priors according to claim 1, wherein said inputting the two-dimensional measurement image and the compressed physical priors vector into a depth-expanded neural network, outputting a high-dimensional spectrum data cube of the object to be observed, comprises: Inputting the two-dimensional measurement image and the compressed physical prior vector into a depth expansion neural network, and carrying out convolution operation on the two-dimensional measurement image and the compressed physical prior vector through the front convolution layer to obtain an initial fusion characteristic; processing the two-dimensional measurement image, the compressed physical prior vector and the initial fusion feature through the cascade module to obtain a target cascade feature; and processing the target cascade characteristics through the tail end convolution layer, and outputting a high-dimensional spectrum data cube of the target to be observed.
4. The method for reconstructing a spectrum data cube based on compressed physical prior according to claim 3, wherein the cascade module comprises M cascade sub-modules connected in sequence, wherein the processing the two-dimensional measurement image, the compressed physical prior vector and the initial fusion feature by the cascade module to obtain a target cascade feature comprises: Inputting the two-dimensional measurement image, the compressed physical prior vector and the initial fusion feature into the cascade module, and processing the two-dimensional measurement image, the compressed physical prior vector and the initial fusion feature through a first cascade sub-module of the cascade module to obtain a first intermediate feature; for the ith cascading sub-module, the following steps are performed: Processing the ith-1 intermediate feature, the two-dimensional measurement image and the compressed physical prior vector output by the last cascade submodule through an ith cascade submodule, and outputting the ith intermediate feature, wherein the value range of i is more than or equal to 2 and less than or equal to M; After being sequentially processed by the M cascading submodules, the M-stage intermediate features output by the M cascading submodules are determined to be target cascading features.
5. The method for reconstructing a spectrum data cube based on a compressed physical prior according to claim 4, wherein the cascade submodule comprises an information fusion unit and a convolution network unit, wherein the processing the i-1 intermediate feature output by the last cascade submodule, the two-dimensional measurement image and the compressed physical prior vector by the i-th cascade submodule, and outputting the i-th intermediate feature comprises: Inputting the i-1 intermediate feature, the two-dimensional measurement image and the compressed physical prior vector into an i-th cascade submodule, and performing channel splicing and convolution fusion on the i-1 intermediate feature, the two-dimensional measurement image and the compressed physical prior vector through the information fusion unit to obtain an i-th multisource fusion feature; And extracting depth features of the ith multisource fusion feature through the convolution network unit, and outputting an ith intermediate feature.
6. A spectroscopic data cube reconstruction apparatus based on compressed physical priors, comprising: the acquisition module is used for acquiring a two-dimensional measurement image formed by code modulation of a target to be observed through the spectrum imaging chip; The acquisition module is also used for acquiring an original transmission spectrum matrix corresponding to each pixel in the spectrum imaging chip; the analysis module is used for carrying out principal component analysis on the original transmission spectrum matrix in the spectrum dimension and extracting principal component feature vectors through a singular value decomposition method; The projection module is used for projecting the original transmission spectrum matrix into an orthogonal subspace formed by the principal component feature vectors to obtain a compressed physical prior vector; The input module is used for inputting the two-dimensional measurement image and the compressed physical prior vector into a depth expansion neural network and outputting a high-dimensional spectrum data cube of the object to be observed, wherein the depth expansion neural network comprises a front convolution layer, a cascade module and a tail end convolution layer; the analysis module is used for: Expanding the original transmission spectrum matrix according to the dimension of the space pixels to obtain a two-dimensional matrix, wherein each row vector of the two-dimensional matrix corresponds to the spectral transmission response data of the space pixels in the spectral imaging chip; Performing decentration pretreatment of principal component analysis on the two-dimensional matrix in a spectrum dimension to obtain a decentration matrix; Calculating a covariance matrix of a spectrum dimension according to the decentration matrix; Singular value decomposition is carried out on the covariance matrix to obtain a singular value matrix; And ordering all the eigenvectors in the singular value matrix in descending order according to singular values to obtain an eigenvector sequence, and extracting principal component eigenvectors from the eigenvector sequence, wherein the descending order ordering rearranges the eigenvectors of the spectrum dimension from high to low according to the information contribution degree of the eigenvectors.
7. An electronic device comprising a memory, a processor and a computer program stored on the memory and running on the processor, wherein the processor implements the method of compressed physical prior-based spectral data cube reconstruction according to any one of claims 1 to 5 when the computer program is executed.
8. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor implements a method of reconstruction of a spectroscopic data cube based on compressed physical priors as claimed in any one of claims 1 to 5.
9. A computer program product comprising a computer program which, when executed by a processor, implements a method of reconstruction of a spectroscopic data cube based on compressed physical priors as claimed in any one of claims 1 to 5.

Description

Method, device, equipment, storage medium and program product for reconstructing spectrum data cube based on compressed physical prior Technical Field The invention relates to the technical field of spectrum reconstruction, in particular to a method, a device, equipment, a storage medium and a program product for reconstructing a spectrum data cube based on compressed physical priori. Background The micro spectral imaging technology is an emerging technology capable of synchronously acquiring the shape and the spectral characteristics of a target space, and is widely applied to a plurality of fields such as environmental monitoring, medical diagnosis, industrial detection, security monitoring and the like by virtue of the advantages of small size, high integration level, strong portability and the like. The core implementation logic of the technology is that a target optical signal incident to a chip is subjected to spectrum coding modulation through a micro-nano structure or an optical filter array integrated on a micro-spectrum imaging chip, then a two-dimensional measurement image formed after coding modulation is acquired by an image sensor, and finally the two-dimensional measurement image is subjected to resolving processing through a specific algorithm, so that a spectrum data cube containing a space dimension-spectrum dimension is recovered. However, conventional spectral reconstruction algorithms typically use a complete high-dimensional transmission spectral matrix to participate in the process of resolving the spectral data cube, which can result in a large amount of network parameters, thereby affecting the reconstruction efficiency of the spectral data cube. Disclosure of Invention The invention provides a method, a device, equipment, a storage medium and a program product for reconstructing a spectrum data cube based on compressed physical priori, which are used for solving the technical problem that the reconstruction efficiency of the spectrum data cube is affected due to the huge network parameters in the traditional spectrum reconstruction algorithm in the prior art. The invention provides a spectrum data cube reconstruction method based on compressed physical priori, comprising the following steps: Acquiring a two-dimensional measurement image formed by code modulation of a target to be observed through a spectrum imaging chip; Acquiring an original transmission spectrum matrix corresponding to each pixel in the spectrum imaging chip; Performing principal component analysis on the original transmission spectrum matrix in a spectrum dimension, and extracting principal component feature vectors through a singular value decomposition method; Projecting the original transmission spectrum matrix into an orthogonal subspace formed by the principal component feature vectors to obtain a compressed physical prior vector; And inputting the two-dimensional measurement image and the compressed physical prior vector into a depth expansion neural network, and outputting a high-dimensional spectrum data cube of the target to be observed. According to the method for reconstructing the spectrum data cube based on the compressed physical prior, the primary component analysis is carried out on the original transmission spectrum matrix in the spectrum dimension, and the feature vector of the primary component is extracted through a singular value decomposition method, which comprises the following steps: Expanding the original transmission spectrum matrix according to the dimension of the space pixels to obtain a two-dimensional matrix, wherein each row vector of the two-dimensional matrix corresponds to the spectral transmission response data of the space pixels in the spectral imaging chip; Performing decentration pretreatment of principal component analysis on the two-dimensional matrix in a spectrum dimension to obtain a decentration matrix; Calculating a covariance matrix of a spectrum dimension according to the decentration matrix; Singular value decomposition is carried out on the covariance matrix to obtain a singular value matrix; and ordering all eigenvectors in the singular value matrix in descending order according to singular values to obtain an eigenvector sequence, and extracting principal component eigenvectors from the eigenvector sequence. According to the method for reconstructing the spectrum data cube based on the compressed physical prior, the method for reconstructing the spectrum data cube based on the compressed physical prior provided by the invention, before inputting the two-dimensional measurement image and the compressed physical prior vector into a deep expansion neural network and outputting the high-dimensional spectrum data cube of the object to be observed, further comprises: Constructing a training sample set of an initial depth expansion neural network, wherein the training sample set comprises a plurality of groups of training samples, and each group of training samples