CN-121982368-A - Real and false target rapid intelligent detection and identification method based on computational spectrum imaging

CN121982368ACN 121982368 ACN121982368 ACN 121982368ACN-121982368-A

Abstract

The invention designs a method for rapidly and intelligently detecting and identifying true and false targets based on computational spectrum imaging, which has strong camouflage property of inflated true and false targets, is difficult to effectively identify by the traditional optical detection means, and is inconvenient for the expansion of business work. The method expands the dimension of the target information to the space dimension and the spectrum dimension, thereby enhancing the distinguishing property of the target, the background and the interference and improving the anti-interference and identification capability. According to the method, target map information is quickly acquired through an imaging system based on a microlens array, then, the image is subjected to spectral reconstruction based on a depth expansion transducer, single-frame measurement image data are restored to be multispectral cubes, a spectral image super-resolution network training technology based on similarity measurement and nucleation self-attention is constructed, the resolution of the spectral image is improved, then, a convolutional neural network encoder and a self-attention mechanism are used for extracting features of spectral spectrum samples, a multi-head self-attention mechanism is constructed to capture complex relations among the samples, and finally, the purpose of accurately detecting true and false inflatable targets is achieved.

Inventors

LIU LUFANG
REN JINLEI
CAI JINGKUN
GENG KEDA
WU TAIHUI
Ji Xiaochuang
GU YUE
GUO JIAWEI
CHAO LUJING
LU YING
ZHU HU
LI LUOGANG
HOU JUNCHEN
XU GUOXIA
YAN YAN
WANG ZHENYA

Assignees

中国航天科技创新研究院

Dates

Publication Date: 20260505
Application Date: 20251224

Claims (10)

1. A real and false target rapid intelligent detection and identification method based on computational spectrum imaging is characterized by comprising the following steps: (1) The parallelized spectrum acquisition and imaging are realized through the design of an imaging system, the snapshot type calculation spectrum imaging is realized, and a target single-frame multispectral image Y is obtained; (2) Performing spectral reconstruction based on a depth expansion transducer on the spectral image Y obtained in the step (1), and reconstructing a high-dimensional spectral image X of a microlens array multichannel from a single-frame multispectral compression measurement image; (3) Establishing a spectrum image super-resolution model for solving according to the reconstructed high-dimensional spectrum image X in the step (2) by combining the spectrum correlation coefficient and the nucleated self-attention to obtain a high-resolution multi-spectrum image X max ; (4) For the high-resolution multispectral image X max , determining a final optimal band subset through a multi-target band selection model As a band spectrum sample; (5) By sampling selected bands of spectrum And extracting features, constructing a true and false target detection model based on a convolutional neural network classifier, and realizing accurate detection of the true and false targets.
2. The method for rapidly and intelligently detecting and identifying true and false targets based on computational spectrum imaging according to claim 1, wherein the step (1) realizes parallelized spectrum acquisition and imaging by designing an imaging system, realizes snapshot type computational spectrum imaging, and is specifically as follows: an imaging system is built, and comprises a front-mounted telescopic system, a view field diaphragm, a collimating lens, a micro lens array, an array filter and a detector; The method comprises the steps that optical information carried by a target is collected through a front-mounted telescopic system, light beams are converged through a view field diaphragm, the light beams are irradiated on a micro lens array in parallel through a collimating lens, then focusing and filtering are completed through the micro lens array and an array filter respectively, high-dimensional spectrum data are compressed into a low-dimensional measurement space during imaging, and finally complete space image information of the target under corresponding wavelengths is obtained in each channel of a detector respectively, namely a single-frame multispectral image Y of the target.
3. The method for rapidly and intelligently detecting and identifying true and false targets based on computed spectrum imaging according to claim 2, wherein the micro-lens array disperses light into a plurality of channels, each micro-lens is responsible for focusing light of a sub-field, and the complex light is separated into monochromatic light through the array filter, so that spectrum information of scene targets is finally obtained.
4. The method for rapidly and intelligently detecting and identifying true and false targets based on computed spectrum imaging according to claim 1, wherein the step (2) is characterized in that the obtained single-frame spectrum image Y is subjected to spectrum reconstruction based on a depth expansion transducer, and a high-dimensional spectrum image X of a microlens array multichannel is reconstructed from a single-frame multispectral compression measurement image, specifically comprising the following steps: (2.1) efficient modeling of spectral features through non-local spectral attention mechanisms, to be achieved Flattened into a space-spectrum sequence N=hw, where C is the number of spectral channels, H, W is the spatial dimension, N represents the total number of feature positions; (2.2) Using the matrix to apply Respectively mapping into a query vector Q, a key vector K and a value vector V, wherein the formula is as follows: Wherein the method comprises the steps of Is a matrix of parameters that can be learned; (2.3) obtaining an attention weight matrix A through Softmax normalization, wherein the formula is as follows: wherein d is the feature channel dimension; (2.4) weighting and aggregating the vector V by using the attention weight matrix A, and finally outputting the characteristic representation of the single-frame spectrum image Y fused with the global information The formula is: (2.5) representation according to characteristics Problem decomposition is performed by using semi-definite quadratic division to solve a low-resolution high-dimensional spectrum image X, U as an auxiliary variable, and the formula is as follows: alternately solving two sub-problems: Data subproblems: a priori sub-problem: Wherein, the Representing the linear mapping operator of the imaging system, In order to regularize the term a priori, As a parameter of the weight-bearing element, For penalty parameters, U is an auxiliary variable, Indicating the optimal solution for X and U obtained under the given conditions.
5. The method for rapidly and intelligently detecting and identifying true and false targets based on computed spectrum imaging according to claim 1, wherein the step (3) is characterized in that according to the reconstructed low-resolution high-dimensional spectrum image X, a spectrum image super-resolution model is established by combining a spectrum correlation coefficient and a nucleation self-attention to solve, and a high-resolution multispectral image X max is obtained, specifically: (3.1) flattening the Low resolution high dimensional spectral image X into a space-spectral sequence N=hw, where C is the number of spectral channels, H, W is the spatial dimension, N represents the total number of feature positions; (3.2) Using the matrix to apply The values are mapped into a query vector Q h , a key vector K h and a value vector V h respectively, and the formulas are as follows: Wherein the method comprises the steps of Is a matrix of parameters that can be learned; (3.3) carrying out the mean value removal and normalization operation on the query vector Q h and the key vector K h , eliminating the vector amplitude difference and only keeping the spectrum curve form; Removing the average value: , Normalization: , Wherein B is the number of spectral bands, And The lengths of vectors Q h and K h , respectively; (3.4) normalizing , And (3) performing kernel function fusion: Due to , More prone to highlight spectral curve morphology similarity in the high-dimensional space, low-resolution multispectral image Is characterized by (a) self-attention Can be expressed as: (3.5) training the model, for a high resolution multispectral image The tensor loop decomposition, which can be expressed as three nuclear tensors, is broad, high, spectral, with the formula: Wherein, the For a nuclear tensor, R is a tensor loop operation; Its low resolution image From high-resolution multispectral images The space sampling matrix P 1 、P 2 and the spectrum sampling matrix P 3 of the calculated spectrum imaging sensor are subjected to dimension reduction, and , is expressed as the following formula: a high-resolution computed spectral image Z, Training to obtain a model loss function as Wherein, the For pre-selected parameters, for setting the weight of each image, F is the Frobenius norm, and the loss function gradient values are used to alternatively optimize the solution Minimizing the loss value, namely obtaining the optimal solution of the model; (3.6) self-attention feature of Low resolution multispectral image X And carrying out the method into a trained model, and solving a high-resolution multispectral image X max .
6. The method for rapid intelligent detection and identification of true and false targets based on computed-spectrum imaging according to claim 1, wherein said step (4) determines final optimal band subset for high-resolution multispectral image X max through a multi-target band selection model As band spectrum samples, specifically: defining a multi-objective band selection problem as a multi-objective optimization problem, the multi-objective optimization problem being formalized as: , Wherein, the Representing an information enrichment target, a redundancy minimization target and a target background separability maximization target respectively; Is a feasible solution space for band selection, each solution Representing a set of bands selected from the original multispectral data, simultaneously seeking a subset of bands that perform well on all three targets, i.e., determining a final optimal subset of bands, by the above-described multi-target form As a band spectrum sample.
7. The method for rapid intelligent detection and identification of true and false targets based on computed tomography according to claim 1, wherein the step (5) is performed by sampling the spectrum of the selected wavelength band Performing feature extraction, constructing a true and false target detection model based on a convolutional neural network classifier, and realizing accurate detection of the true and false targets, wherein the method specifically comprises the following steps: (5.1) extracting features, namely extracting features of band spectrum samples by using a convolutional neural network and a self-attention mechanism, automatically filtering noise samples with low correlation with other samples, and processing the extracted features by using a multi-layer perceptron end in a multi-head self-attention mechanism, wherein all linear layers of the neural network structure use a plurality of units and a ReLU activation function; (5.2) training a true and false target detection model, comprising the following steps of (A) Data preparation, namely constructing a spectrum data set, wherein the data set comprises real and false target samples; (b) The loss function is designed according to the loss and precision change curve, and the difference between the model output and the real label is measured by selecting the cross entropy loss function; (c) Training the real and false target detection model by using Adadelta optimizers, updating model parameters according to gradients of the loss function, and achieving the goal of minimizing the loss function so as to improve the classification performance of the real and false target detection model on real and false targets; and (5.3) detecting the true and false targets of the band spectrum sample features extracted in the step (5.1) by using a trained true and false target detection model.
8. A processor, characterized in that the processor is configured to run a program, wherein the program, when run, performs the method according to any of claims 1-7.
9. A non-volatile storage medium comprising a computer program product which, when executed, performs the method of any of claims 1-7.
10. A computer program product, characterized in that the computer program product comprises a computer program which, when executed by a processor, implements the steps of the method according to any one of claims 1-7.

Description

Real and false target rapid intelligent detection and identification method based on computational spectrum imaging Technical Field The invention relates to a real and false target rapid intelligent detection and identification method based on computational spectrum imaging, and belongs to the technical field of target detection and identification. Background Along with the development of science and technology, camouflage technology is developing towards multiband, diversified and intelligent, the diversity and complexity presented by the camouflage technology provide great challenges for an optical detector, and how to ensure that ground real and false targets can be detected and identified robustly and rapidly becomes a urgent problem to be solved. The common detection method of the true and false targets mainly comprises a true and false target identification technology based on infrared imaging, a true and false target identification technology based on polarized light imaging and the like. With the development of spectroscopic techniques, hyperspectral imaging techniques are widely used in the field of remote sensing. The spectrum imaging technology is a front remote sensing technology for simultaneously acquiring two-dimensional space target spectrum information, can simultaneously image a plurality of wave bands, and the light reflectivity of images of specific wave bands is different for different materials. The technology is beneficial to accurately detecting and identifying the specified target from the complex background and is beneficial to improving the real-time performance and accuracy of target detection, identification and classification. However, although the spectrum target recognition tracking technology has higher spatial resolution and spectrum resolution, and can acquire the fine information of the target, the problems of complex spectrum imaging system, large weight, large spectrum data volume, difficult real-time response, high spectrum noise pollution, high recognition difficulty, low algorithm accuracy, poor robustness and the like exist, and the real and false targets are difficult to be rapidly detected and identified in a complex environment, so that further research on a method for rapidly detecting and identifying the real and false targets is necessary. Disclosure of Invention The technical problem to be solved is to overcome the defects of the prior art, and provide a real and false target rapid intelligent detection and identification method based on computational spectrum imaging, so that the real and false targets can be rapidly, efficiently and accurately identified. The technical scheme of the invention is as follows: A real and false target rapid intelligent detection and identification method based on computational spectrum imaging comprises the following steps: (1) The parallelized spectrum acquisition and imaging are realized through the design of an imaging system, the snapshot type calculation spectrum imaging is realized, and a target single-frame multispectral image Y is obtained; (2) Performing spectral reconstruction based on a depth expansion transducer on the spectral image Y obtained in the step (1), and reconstructing a high-dimensional spectral image X of a microlens array multichannel from a single-frame multispectral compression measurement image; (3) Establishing a spectrum image super-resolution model for solving according to the reconstructed high-dimensional spectrum image X in the step (2) by combining the spectrum correlation coefficient and the nucleated self-attention to obtain a high-resolution multi-spectrum image X max; (4) For the high-resolution multispectral image X max, determining a final optimal band subset through a multi-target band selection model As a band spectrum sample; (5) By sampling selected bands of spectrum And extracting features, constructing a true and false target detection model based on a convolutional neural network classifier, and realizing accurate detection of the true and false targets. Furthermore, the step (1) realizes parallelized spectrum acquisition and imaging by designing an imaging system, and realizes snapshot type calculation spectrum imaging, which specifically comprises the following steps: an imaging system is built, and comprises a front-mounted telescopic system, a view field diaphragm, a collimating lens, a micro lens array, an array filter and a detector; The method comprises the steps that optical information carried by a target is collected through a front-mounted telescopic system, light beams are converged through a view field diaphragm, the light beams are irradiated on a micro lens array in parallel through a collimating lens, then focusing and filtering are completed through the micro lens array and an array filter respectively, high-dimensional spectrum data are compressed into a low-dimensional measurement space during imaging, and finally complete space image information of the target