Point spread function optimization method of multispectral compression imaging system

Abstract

The invention discloses a point spread function optimization method for a multispectral compression imaging system, and the method comprises the steps: constructing two optimization objective functions of the multispectral compression imaging system, and proposing a corresponding multi-objective optimization algorithm; wherein one objective function is used for evaluating light transmission performance, and the other objective function is used for evaluating cross-correlation characteristics among multiple spectrums; the optimization algorithm is based on an NSGAII algorithm, and compared with a traditional genetic algorithm, records of dominating times are introduced, so that the calculation complexity of a sorting process is effectively reduced; an elite strategy is introduced, a screening mechanism is introduced for individuals participating in heredity, and the heredity process is optimized. And the optimized system point spread function can transmit more perfect space and spectrum information to a greater extent, so that the subsequent reconstruction precision can be improved.

Description

Technical Field

[0001] The invention relates to the field of compression imaging in computational optics, and in particular to a point spread function optimization method for a multi-spectral compression imaging system. Background Art

[0002] Spectral imaging can provide images with rich spatial and spectral information. It is an emerging scientific and engineering application tool with wide application value in many fields. Traditional spectral imaging modes rely on spatial scanning or spectral scanning, and the scanning mechanism is relatively complex. For example, the swing-scanning spectral imaging system images one row or a small area at a time. The spectral information is obtained through the spectroscopic component, and the full field of view information needs to be obtained through the scanning component and the platform movement, and the field of view needs to be stitched. Scanning spectral imaging can achieve high spatial and spectral resolution, but the temporal resolution is low, and it is not suitable for scenes with high real-time requirements. Each sub-aperture of the multi-aperture spectral imaging system images a specific band separately, with high spatial and temporal resolution, but the number of bands is limited.

[0003] In order to solve the above constraints between spatial resolution, temporal resolution and spectral resolution, computational spectral imaging technology came into being. This technology uses a specially designed optical system to encode and compress the multi-spectral image information of the scene, which is received by a wide-band detector, and then reconstructs a three-dimensional spectral image from the two-dimensional modulated image through an image reconstruction algorithm. In this way, the requirements for the hardware of the imaging system are mostly transformed into requirements for the algorithm, which can achieve lightweight and miniaturization of the optical system and reduce the amount of data stored and transmitted for the image.

[0004] How to design an appropriate optical system to modulate three-dimensional spatial-spectral information is one of the key technologies for realizing computational spectral imaging. At present, multi-spectral information modulation is mainly achieved through three ways, namely image plane coding, spectral coding and point spread function engineering (Guo Jiaqi, Fan Benxuan, Liu Xin, et al. Computational spectral imaging: light field coding and algorithm decoding [J]. Laser and Optoelectronics Progress, 2024, 61 (16).). Image plane encoding is to image the object to be tested onto the intermediate image plane, compress and encode the intermediate image to obtain a compressed image. The encoding function is generally a binary function. Redundant information is eliminated by blocking pixels, and spectral information is reused in space. Then the image is imaged to the detector to reconstruct the image to be tested. The principle of the spectral response encoding method is to form a spectral filter array (SFA) through narrowband filtering or specially designed multi-spectral response, and then reconstruct the spectral image according to the complementary information provided by different channels. The key to this method is the design of the spectral response function (SRF). The principle of point spread function (PSF) encoding is to change the PSF of the system through additional optical elements as encoding devices, implement encoding using the characteristics of PSF changing with wavelength, and then reconstruct the multi-spectral data cube. PSF encoding is mainly based on wave optics. In physical implementation, a diffractive optical element (DOE) with a certain microstructure is generally used to change the amplitude or phase distribution of the incident light wave, thereby obtaining a PSF that is conducive to spectral reconstruction.

[0005] Compared with image plane coding, which is only based on ideal geometric optics for analysis, it is also necessary to calibrate the reconstruction algorithm according to the aberration in practical application. The aberration caused by PSF in PSF coding, especially the chromatic aberration, can directly become the basis for spectral image reconstruction; compared with the spectral response method, although the spectral response method encoding device SFA is also very light and thin, it is difficult to manufacture, and the design and manufacture of the DOE element of the PSF encoding device is more mature and reliable. Therefore, the advantages of PSF coding are outstanding, and there is no need to realize the optical elements of explicit coding. The optical system is simple and compact, and the miniaturization and lightweight of the system can be realized, and even the existing imaging equipment used in daily life can be directly transformed into a spectral imaging equipment. The present invention is a related technical research on PSF coding.

[0006] For multi-spectral compression imaging systems, the quality of the reconstructed image directly affects the application effect. However, there is still a gap between the reconstruction effect and the quality of direct acquisition, which has become a key factor restricting its further application. The quality of the reconstructed image is closely related to the measurement matrix in the compression imaging process. A good measurement matrix can transmit more information flux in the sampling and reconstruction process, thereby having a higher reconstructed image quality. The measurement matrix corresponds to the structural design of the encoding device in PSF coding. At present, most of them are randomly generated based on the theory of compressed sensing (CS), or are manually designed. The measurement matrix can be divided into a deterministic measurement matrix and a random measurement matrix (Tong Qi. Research on key technologies of infrared compression imaging [D]. Harbin Engineering University [2023-11-06].). The main methods for generating deterministic measurement matrices include polynomial measurement matrix, block polynomial measurement matrix, algebraic curve measurement matrix, etc. Its advantages are that it can save storage space through design, and it can also design corresponding reconstruction algorithms to improve the speed and accuracy of reconstruction; its disadvantages are that it can only be designed for certain specific scenes, and it has high requirements on the sparsity of the signal. It is extremely difficult to design for highly complex signals such as natural scenes, and the corresponding reconstruction success rate is also low. In contrast, the advantage of the random measurement matrix is ​​that it does not require high sparsity of the target signal and is more adaptable to various natural scenes. Usually, the measurement matrix constructed directly is often not optimal. Therefore, it is very necessary to optimize the random measurement matrix and thus optimize the PSF to improve the reconstruction quality. This optimization process is also called PSF engineering (Jia S, Vaughan JC, Zhuang X. Isotropic 3D Super-resolution Imaging with a Self-bending Point Spread Function [J]. Nature Photonics, 2014, 8 (4): 302-306. DOI: 10.1038 / nphoton.2014.13.).

[0007] At present, the commonly used optimization algorithms in PSF engineering include genetic algorithms, ant colony algorithms, etc. (Shechtman Y, Sahl SJ, Backer AS, et al. Optimal point spread function design for 3D imaging. [J]. Physical Review Letters, 2014, 113 (13): 133902.), (Conkey DB, Brown AN, Caravaca-Aguirre AM, et al. Genetic algorithm optimization for focusing through turbid media in noisy environments [J]. Optics express, 2012, 20 (5): 4840-4849.), and the optimization target needs to design corresponding objective functions for different application scenarios, usually optimizing indicators such as uniformity and clarity of the light spot (Feng, Fan, Yang, et al. Multi-objective optimization genetic algorithm for multi-point light focusing in wavefront shaping. [J]. Optics express, 2019, 27(25): 36459-36473.), and for the application field of measurement matrix optimization in multi-spectral compression imaging system, the optimization target and optimization algorithm of PSF engineering need to be reasonably designed according to application requirements. The present invention focuses on the research on the optimization design method of PSF engineering for multi-spectral compression imaging. Summary of the invention

[0008] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a point spread function optimization method for a multi-spectral compression imaging system, that is, to optimize the system PSF by optimizing the design of the compression sampling matrix through an optimization algorithm, so that the optical transmission information flux is maximized and the correlation between multiple spectral bands is minimized, while increasing the imaging spatial resolution and spectral resolution, thereby improving the reconstruction accuracy.

[0009] In order to achieve the purpose of the present invention, the present invention adopts the following technical solutions:

[0010] A point spread function optimization method for a multi-spectral compression imaging system comprises the following steps:

[0011] (1) Generate N initial random masks as the initial population and calculate the objective function (f1, f2) of each mask; where the first indicator f1 is the double integral of the modulation transfer function in the frequency domain, and the second indicator f2 is the condition number of the correlation matrix of the point spread function between spectra;

[0012] (2) According to the optimization objective function value (f1, f2) of each mask, the hierarchy and crowding distance are determined according to the non-dominated relationship, and a fast non-dominated sort is performed; four child phase templates are generated; the selection, crossover, and mutation operations are repeated N / 4 times to obtain the first generation of N child phase templates, and the objective function (f1, f2) of each mask is calculated;

[0013] (3) Mixing N parent phase templates with N offspring phase templates to obtain 2N phase templates; performing fast non-dominated sorting and crowding distance calculation according to the optimization objective function value (f1, f2) of each mask; retaining the first N offspring masks with lower level numbers and larger crowding distances from the 2N mixed masks as new offspring masks to obtain a new population;

[0014] (4) Determine whether the current population has reached the number of cycles; if so, output the N masks of the current population as the optimal solution; otherwise, return to step (2) and repeat steps (2) to (3) for the new population; the final output result is the optimal population, and the first ranked one is the optimal phase template.

[0015] Furthermore, in step (1), the calculation formula of f1 is:

[0016]

[0017] Among them, λ mid is the median spectrum of the multi-spectral bands, (u,v) is the frequency domain coordinate; f1 is used to evaluate the light transmission performance.

[0018] Furthermore, in step (1), the calculation formula of f2 is:

[0019] f2=Cond(Corr),

[0020]

[0021]

[0022] Among them, Cond represents the condition number of the matrix, Corr is a matrix composed of the normalized mutual correlation coefficients between the PSFs of each spectral segment; the symbol ☆ represents the cross-correlation operation; f2 is used to evaluate the cross-correlation characteristics between multiple spectra.

[0023] Furthermore, the objective function (f1, f2) formula is:

[0024] min{f1,f2}.

[0025] Furthermore, in step (2), the operations of selection, crossover and mutation are specifically as follows:

[0026] Selection: In the current population P, according to the principle of the smallest level of the optimal frontier of non-dominated sorting and the smallest crowding distance within the layer, four phase template masks are selected as a group according to the sorting results, and the mask pairs of the two groups of parent phase templates are determined according to the four-element tournament principle;

[0027] Crossover: Each pair of parent phase templates is randomly and uniformly crossed with a probability of 0 to 1;

[0028] Mutation: All masks are randomly inverted with a probability of 0 to 1.

[0029] Furthermore, the point spread function of the multi-spectral compression imaging system is optimized using the NSGAII algorithm, wherein the main framework adopts the GA genetic algorithm.

[0030] Furthermore, each phase template mask i Corresponding to a body element x i ,In order to make each objective function have an equivalent influence on the crowding distance, f1 and f2 are normalized respectively during the calculation process.

[0031] Furthermore, the types of the phase template include superpixel random template and Gaussian noise template; the corresponding optimization parameters are the Gaussian mean variance of the superpixel template pixel value and the Gaussian noise template.

[0032] Furthermore, the multi-spectral compression imaging system uses a collimator as a short-wave infrared target source, and by replacing a graticule and a Polo plate as different targets, the emitted parallel light is focused on the object plane of the first aperture through a first imaging lens group with a focal length of f, forming an object plane target source, and enters a 4f system. In the 4f system, polarization and filtering are first performed by a polarizer and a filter wheel to obtain polarized light of a selected band. A collimator lens is used as the first lens of the 4f system to realize Fourier frequency domain transformation of the target, and frequency domain modulation is realized by a random phase on a spatial light modulator through a beam splitter. A pair of orthogonal cylindrical lens groups are used as the second lens of the 4f system to realize Fourier inverse transformation, and finally the modulated image is captured by a camera at the image plane.

[0033] Further, the multispectral compression imaging process of the multispectral compression imaging system includes point spread function calibration and multispectral imaging;

[0034] During the point spread function calibration process, the camera is set at the midpoint of the double focus of the cylindrical lens to obtain the best point spread function speckle; the regular spots are disrupted by random modulation of the plate phase of the spatial light modulator on the Fourier plane, and the resulting fine structure is regarded as a smooth speckle pattern; by controlling the filter wheel, the point spread function speckle in 6 spectra is recorded in turn to complete the point spread function calibration;

[0035] In the multispectral imaging process, real objects are used to replace point sources to achieve multispectral compression imaging; when the corrected multispectral point spread function and compressed image are available, the multispectral image is reconstructed by solving the ill-posed inverse problem with sparsity constraints.

[0036] The beneficial effects of the present invention are as follows:

[0037] 1. The present invention constructs two key indicators of the multi-spectral compression imaging system, one to evaluate the light transfer performance and the other to evaluate the cross-correlation characteristics between multiple spectra. The designed optimization target can effectively increase the spatial and spectral information flux of the system and help improve the reconstruction accuracy.

[0038] 2. The present invention optimizes the PSF optimization performance by using the NSGAII algorithm for multi-objective optimization problems. In the process of non-dominated sorting, the record of the number of dominance is introduced, and the crowding distance is introduced to increase the priority within the layer, further improving the performance of the optimization non-dominated sorting step and effectively reducing the computational complexity; the elite strategy is introduced, and a screening mechanism is introduced for the individuals participating in inheritance to optimize the inheritance process. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the working principle of the single-exposure multi-spectral compression imaging system of the present invention;

[0040] Figure 2 Schematic diagram of random phase template distribution of the present invention (superpixel size 60);

[0041] Figure 3 It is a schematic diagram of the point spread function calibrated within the working spectrum range of 1510nm to 1560nm of the present invention;

[0042] Figure 4 Flow chart of the NSGAII algorithm for PSF optimization of the multi-spectral compression imaging system of the present invention. DETAILED DESCRIPTION

[0043] In order to better understand the technical solution of the present application, the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0044] It should be clear that the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0045] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0046] The multispectral imaging optical device used in multispectral imaging is as follows: Figure 1 As shown. A collimator is used as the short-wave infrared imaging (SWIR) target source. By replacing the graticule and the Polo plate as different targets, the outgoing parallel light is focused on the aperture 1 (object plane) through the first imaging lens group with a focal length of f = 200 mm, forming an object plane target source, and then enters the 4f system (a typical filtering system in the optical field, composed of two Fourier lenses with a focal length of f. The distance from the system object plane to the first lens, the distance from the first lens to the second lens, and the distance from the second lens to the image plane are single focal length f, double focal length 2f, and single focal length f, respectively). In the 4f system, the polarizer and filter wheel are first used for polarization and filtering to obtain polarized light of the selected wavelength band. The collimator lens is used as the first lens group of the 4f system to realize the Fourier frequency domain transformation of the target, and is then transmitted to the spatial light modulator (Spatial light modulator) through the beam splitter. The random phase on the SLM on the Fourier plane is used to realize the modulation in the frequency domain. A pair of orthogonal cylindrical lens groups are used as the second lens of the 4f system to realize the inverse Fourier transform. The modulated image is finally captured by the SWIR camera at the image plane. In this process, the astigmatism of the cylindrical lens ensures the distinguishability of different spectral bands. The specific multispectral compression imaging process includes two steps: PSF calibration and multispectral imaging. In the PSF calibration step, the camera is set at the midpoint of the double focus of the cylindrical lens to obtain the best PSF speckle. The random modulation of the flat plate phase of the SLM on the Fourier plane disrupts the regular spots and produces many fine structures, which can be regarded as smooth speckle patterns. The PSF calibration can be completed by controlling the filter wheel and recording the PSF speckle in 6 spectra in sequence. In the multispectral imaging step, the point source is replaced by a real object to realize multispectral compression imaging. When the corrected multispectral PSF and compressed image are available, the multispectral image can be reconstructed by solving the ill-posed inverse problem with sparsity constraints.

[0047] The PSF optimization process described in the present invention is performed before the PSF calibration step, and the specific design is as follows:

[0048] 1. First, we need to select the optimization target of PSF optimization design.

[0049] On the one hand, it is hoped that the PSF obtained through measurement matrix coding can achieve good image quality in a general sense and preserve as much detail information as possible. Indicators such as image variance, image information entropy, and optical transfer function can be used. Here, the present invention chooses to use the optical transfer function to measure the information flux of the imaging system. Specifically, the modulus of the optical transfer function, that is, the modulation transfer function (MTF), calculates the double integral in the frequency domain space as an evaluation index of the system information flux:

[0050]

[0051] Among them, λ mid is the median spectrum of the multi-spectral segments, (u, v) is the frequency domain coordinate, and takes a negative value to ensure that the optimization target is the minimum value. In order to reduce the influence of high-frequency noise in the frequency domain space, the present invention only selects the central area as the integration range instead of the entire frequency domain space.

[0052] On the other hand, for multi-spectral compression imaging, the present invention hopes to make the PSF between spectral bands have greater differentiation while taking into account the imaging quality, thereby helping to improve the spectral resolution of the reconstructed image. This characteristic can be quantitatively measured by the condition number of the mutual correlation coefficient matrix of the PSF between spectral bands, and the formula is as follows:

[0053] f2=Cond(Corr), (2)

[0054] Among them, Cond represents the condition number of the matrix, Corr is a matrix composed of the normalized mutual correlation coefficients between the PSFs of each spectrum segment, and each matrix element is calculated according to the following steps: First, the mutual correlation coefficients between each PSF are calculated:

[0055]

[0056] The symbol ☆ here represents the cross-correlation operation. In order to eliminate the difference in absolute amplitude between the coefficients, the diagonal elements are used for normalization, and the result is:

[0057]

[0058] The smaller the element of the correlation matrix, the smaller the condition number, which means the greater the difference between the related spectrum segments, and the better the PSF. In particular, when the element of the correlation matrix is ​​zero, the correlation matrix is ​​a diagonal matrix, and the minimum condition number is 1. At this time, the spectrum segments are not correlated, that is, the difference between the spectrum segments is the largest.

[0059] Therefore, the optimization objective formula can be expressed as formula (5):

[0060] min{f1,f2} (5)

[0061] 2. Secondly, we need to select the optimization algorithm for PSF optimization design.

[0062] According to the above, the number of optimization targets is two, so the optimization design problem of PSF belongs to "multi-objective nonlinear integer programming problem", and there are many optional optimization algorithms. Existing related research on wavefront shaping problem shows that the optimization effect of genetic algorithm performs well, and researchers have also made corresponding improvements and implementations for different specific applications. For the optimization problem with "multi-objective" as the optimization target, NSGAII (non-dominated sorting genetic algorithmII) is one of the most popular genetic algorithms, which is also the algorithm (Deb K, Pratap A, Agarwal S, et al. A fast and elitist multiobjective genetic algorithm: NSGA-II, IEEE Trans. on Evol [J]. IEEE Transactions on Evolutionary Computation, 2002, 6.) that the present invention intends to adopt.

[0063] The main framework of the algorithm follows the GA (genetic algorithm) genetic algorithm, which has three steps: selection, crossover, and mutation. At the same time, compared with the genetic algorithm, there are three key improvements: first, in terms of the processing of multiple optimization objectives, the non-dominated relationship is used to generate the Pareto front as the basis for the genetic algorithm template individual selection; second, in the process of non-dominated sorting, the statistics of the number of dominance are introduced to achieve fast sorting of the template individual level, and the crowding distance is introduced to perform intra-layer priority sorting of the template individual. The combination of the two effectively reduces the computational complexity; third, the elite strategy is introduced, and a screening mechanism is introduced for the template individuals participating in inheritance to optimize the inheritance process.

[0064] 3. In the process of optimizing the PSF, each phase template mask i Corresponding to a body element x i In order to make each objective function have an equivalent influence on the crowding distance, f1 and f2 should be normalized separately during the calculation process.

[0065] Finally, the type of the phase template can be a superpixel random template, a Gaussian noise template, etc., and the optimization parameters can be the pixel value of the superpixel template, the Gaussian mean variance of the Gaussian noise template, etc.

[0066] After selecting the optimization objective and optimization algorithm, the specific implementation method of the PSF optimization design process is described as follows:

[0067] If a randomly distributed superpixel phase template is selected as the phase template type used for modulation, then the optimization object is the pixel grayscale value of the template. Taking the 1920*1080 SLM scale as an example, a random phase template with a superpixel size of 60 is Figure 2 As shown. The wavelength is 1510nm ~ 1560nm as the system working spectrum range, and a set of point spread functions is collected every 10nm, a total of 6 point spread functions, as shown Figure 3 The specific algorithm flow of using NSGAII algorithm to realize PSF optimization template design is as follows Figure 4 As shown, the present invention provides a point spread function optimization method for a multi-spectral compression imaging system, and the detailed steps are as follows:

[0068] Step 1: Initial population generation.

[0069] Generate N initial random masks as the initial population P0, and calculate the objective function (f1, f2) of each mask. The first indicator f1 is the double integral of the modulation transfer function (MTF) in the frequency domain, calculated as formula (1); the second indicator f2 is the condition number of the inter-spectral PSF correlation matrix, calculated as follows:

[0070]

[0071] f2=Cond(Corr),

[0072]

[0073]

[0074] The optimization objective formula is as follows:

[0075] min{f1,f2},

[0076] Among them, λ mid is the median spectrum of the multi-spectral segments, (u, v) is the frequency domain coordinate, which takes a negative value to ensure that the optimization goal is to take the minimum value; Cond represents the condition number of the matrix, Corr is a matrix composed of the normalized mutual correlation coefficients between the PSFs of each spectrum segment; the symbol ☆ represents the cross-correlation operation.

[0077] Step 2: Generate offspring individuals.

[0078] According to the optimization objective function value (f1, f2) of each mask, the hierarchy and crowding distance are determined according to the non-dominated relationship, and a fast non-dominated sorting is performed.

[0079] Selection: In the current population P, according to the principle of the smallest level of the optimal frontier of non-dominated sorting and the smallest crowding distance within the layer, four phase template masks are selected as a group according to the sorting results, and they are combined in pairs according to the four-element tournament principle to determine the mask pairs of the two sets of parent phase templates.

[0080] Crossover: Each pair of parent phase templates is randomly and evenly crossed with a certain probability (0-1).

[0081] Mutation: All masks are randomly reversed with a certain probability (0-1).

[0082] Thus, four sub-generation phase templates are generated. Repeat the above steps of "selection, crossover, mutation" N / 4 times to obtain the first generation of N sub-phase templates, and calculate the objective function (f1, f2) of each mask.

[0083] Step 3: Elite retention and new population determination.

[0084] Mix N parent phase templates with N offspring phase templates to obtain 2N phase templates.

[0085] Fast non-dominated sorting and crowding distance calculation are performed according to the optimized objective function value (f1, f2) of each mask.

[0086] From the 2N mixed masks, the first N child masks with lower level numbers and larger crowding distances are retained as new child masks, where the level numbers are given priority, and the crowding distance is considered when the numbers are the same. Thus, a new population is obtained.

[0087] Step 4: End conditional judgment.

[0088] Determine whether the current population meets the loop end condition of "reaching the number of cycles". If it does, output the N masks of the current population as the optimal solution; otherwise, return to step 2 and repeat the above steps 2-3 for the new population. The value of the number of cycles comprehensively considers the optimization effect and optimization time, and there is no fixed range. In this embodiment, the value is 20 to 100 times.

[0089] The final output result is the best population, and the first ranked one is the best phase template, thus completing the PSF optimization.

[0090] The phase template obtained by the optimization result can be used in the subsequent multi-spectral compression imaging process, that is, the process of "PSF calibration-imaging sampling-reconstruction". The optimized system PSF can transmit more complete spatial and spectral information to a greater extent, which is conducive to improving the subsequent reconstruction accuracy.

[0091] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A point spread function optimization method for a multi-spectral compression imaging system, characterized in that: The following steps are involved: (1) Generate N initial random masks as the initial population and calculate the objective function (f1, f2) of each mask; where the first indicator f1 is the double integral of the modulation transfer function in the frequency domain, and the second indicator f2 is the condition number of the correlation matrix of the point spread function between spectra; (2) According to the optimization objective function value (f1, f2) of each mask, the hierarchy and crowding distance are determined according to the non-dominated relationship, and a fast non-dominated sort is performed; four child phase templates are generated; the selection, crossover, and mutation operations are repeated N / 4 times to obtain the first generation of N child phase templates, and the objective function (f1, f2) of each mask is calculated; (3) Mixing N parent phase templates with N offspring phase templates to obtain 2N phase templates; performing fast non-dominated sorting and crowding distance calculation according to the optimization objective function value (f1, f2) of each mask; retaining the first N offspring masks with lower level numbers and larger crowding distances from the 2N mixed masks as new offspring masks to obtain a new population; (4) Determine whether the current population has reached the number of cycles; if so, output the N masks of the current population as the optimal solution; otherwise, return to step (2) and repeat steps (2) to (3) for the new population; the final output result is the optimal population, and the first ranked one is the optimal phase template.

2. The point spread function optimization method of the multi-spectral compression imaging system according to claim 1, characterized in that: In step (1), the calculation formula of f1 is: Among them, λ mid is the median spectrum of the multi-spectral bands, (u,v) is the frequency domain coordinate; f1 is used to evaluate the light transmission performance.

3. The point spread function optimization method of the multi-spectral compression imaging system according to claim 2, characterized in that: In step (1), the calculation formula of f2 is: f2=Cond(Corr), Among them, Cond represents the condition number of the matrix, Corr is a matrix composed of the normalized mutual correlation coefficients between the PSFs of each spectral segment; the symbol ☆ represents the cross-correlation operation; f2 is used to evaluate the cross-correlation characteristics between multiple spectra.

4. The point spread function optimization method of the multi-spectral compression imaging system according to claim 3, characterized in that: The objective function (f1, f2) formula is: min{f1,f2}.

5. The point spread function optimization method of the multi-spectral compression imaging system according to claim 1, characterized in that: In step (2), the operations of selection, crossover and mutation are specifically as follows: Selection: In the current population P, according to the principle of the smallest level of the optimal frontier of non-dominated sorting and the smallest crowding distance within the layer, four phase template masks are selected as a group according to the sorting results, and the mask pairs of the two groups of parent phase templates are determined according to the four-element tournament principle; Crossover: Each pair of parent phase templates is randomly and uniformly crossed with a probability of 0 to 1; Mutation: All masks are randomly inverted with a probability of 0 to 1.

6. The point spread function optimization method of the multi-spectral compression imaging system according to claim 1, characterized in that: The point spread function of the multi-spectral compression imaging system is optimized using the NSGAII algorithm, wherein the main framework adopts the GA genetic algorithm.

7. The point spread function optimization method of the multi-spectral compression imaging system according to claim 1, characterized in that: Each phase template mask i Corresponding to a body element x i ,In order to make each objective function have an equivalent influence on the crowding distance, f1 and f2 are normalized respectively during the calculation process.

8. The point spread function optimization method of the multi-spectral compression imaging system according to claim 1, characterized in that: The types of the phase template include super-pixel random template and Gaussian noise template; the corresponding optimization parameters are the Gaussian mean variance of the super-pixel template pixel value and the Gaussian noise template.

9. The point spread function optimization method of the multi-spectral compression imaging system according to claim 1, characterized in that: The multi-spectral compression imaging system uses a collimator as a short-wave infrared target source, and by replacing a graticule and a Polo plate as different targets, the emitted parallel light is focused on the object plane of the first aperture through a first imaging lens group with a focal length of f, forming an object plane target source, and enters a 4f system. In the 4f system, polarization and filtering are first performed by a polarizer and a filter wheel to obtain polarized light of a selected band. A collimator lens is used as the first lens of the 4f system to realize Fourier frequency domain transformation of the target, and frequency domain modulation is realized by a random phase on a spatial light modulator through a beam splitter. A pair of orthogonal cylindrical lens groups are used as the second lens of the 4f system to realize Fourier inverse transformation, and finally the modulated image is captured by a camera at the image plane.

10. The point spread function optimization method of the multi-spectral compression imaging system according to claim 9, characterized in that: The multi-spectral compression imaging process of the multi-spectral compression imaging system includes point spread function calibration and multi-spectral imaging; During the point spread function calibration process, the camera is set at the midpoint of the double focus of the cylindrical lens to obtain the best point spread function speckle; the regular spots are disrupted by random modulation of the plate phase of the spatial light modulator on the Fourier plane, and the resulting fine structure is regarded as a smooth speckle pattern; by controlling the filter wheel, the point spread function speckle in 6 spectra is recorded in turn to complete the point spread function calibration; In the multispectral imaging process, real objects are used to replace point sources to achieve multispectral compression imaging; when the corrected multispectral point spread function and compressed image are available, the multispectral image is reconstructed by solving the ill-posed inverse problem with sparsity constraints.

Images