A variable resolution compressed sensing hyperspectral computational imaging system and device

Abstract

This invention discloses a variable-resolution compressed sensing hyperspectral computational imaging system and apparatus, comprising: a modulation unit array and a control module; the control module is used to output coded signals to the modulation unit array according to a partitioned identical modulation method; the modulation unit array is used to modulate the light signals emitted by the target scene converging on itself according to the coded signals and then project them onto an area array detector; the partitioned identical modulation method is as follows: according to the actual required image spatial resolution, the modulation unit array is divided into sub-blocks of the same size and non-overlapping; the control module outputs the same coded signal to the modulation units in each sub-block, so that the modulation units in the same sub-block modulate the light signals in the same way, thereby merging all the modulation units in each sub-block into an equivalent modulation unit, realizing the variable modulation resolution of the modulation unit array; using this invention, the flexible variable spatial resolution of hyperspectral imaging can be achieved without changing the hardware structure.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral computational imaging technology, specifically to a variable resolution compressed sensing hyperspectral computational imaging system and device. Background Technology

[0002] Hyperspectral imaging technology can acquire images of objects or substances across dozens or even hundreds of spectral bands. These bands are very narrow and continuous, allowing for accurate depiction of the complete spectral curve of each pixel in the image. With its unique advantages of high spectral resolution and integrated image-spectrum mapping, hyperspectral imaging technology has attracted widespread attention from academia and industry, yielding a series of research results and playing an important role in various fields.

[0003] Due to the superior performance of Digital Micromirror Devices (DMDs) in terms of response speed, modulation accuracy, and uniformity, most hyperspectral imaging systems currently utilize DMDs for spatial dimension encoding. For existing DMD compressed sensing hyperspectral imaging systems, the system's spatial resolution depends on the DMD. This means that once the DMD parameters and the distances between system components are determined, the system's spatial resolution is also fixed. The final image obtained through reconstruction algorithms will have the same resolution as the DMD, making flexible adjustments difficult. Furthermore, at the same compression ratio, higher system resolution requires longer sampling times, resulting in greater computational demands and longer reconstruction times. Due to these limitations, the flexibility and practicality of DMD compressed sensing hyperspectral imaging systems require further improvement. Summary of the Invention

[0004] In view of this, the present invention provides a variable resolution compressed sensing hyperspectral computational imaging system and device, which can achieve flexible and variable spatial resolution of hyperspectral imaging without changing the hardware structure.

[0005] To solve the above-mentioned technical problems, the present invention is implemented as follows:

[0006] A variable resolution compressed sensing hyperspectral computational imaging system includes: a modulation unit array, a control module, an area array detector, and a reconstruction module;

[0007] The control module is used to output coded signals to the modulation unit array according to the partitioned same-value modulation method;

[0008] The modulation unit array is used to modulate the light signal emitted by the target scene that converges on it according to the encoded signal and then project it onto the area array detector; each modulation unit in the modulation unit array independently receives the encoded signal output by the control module;

[0009] The partitioned same-value modulation method is as follows: according to the actual required image spatial resolution, the modulation unit array is divided into sub-blocks of the same size and non-overlapping. The control module outputs the same encoded signal to the modulation unit in each sub-block, so that the modulation unit in the same sub-block modulates the optical signal in the same way, thereby merging all the modulation units in each sub-block into an equivalent modulation unit, so that the modulation resolution of the modulation unit array can be variable.

[0010] The array detector is used to sample the received optical signal and input the sampling result into the reconstruction module;

[0011] The reconstruction module is used to reconstruct a hyperspectral image of the target scene at the modulation resolution based on the sampling results.

[0012] Preferably, it further includes a super-resolution reconstruction module, used to obtain a higher resolution image of the target scene from the hyperspectral image of the target scene at the modulation resolution obtained by the reconstruction module through a compressed sensing reconstruction algorithm, thereby achieving super-resolution reconstruction.

[0013] Preferably, the partitioned in-line modulation method includes the following steps:

[0014] Step 1: The number of detector pixel columns of the area array detector is M. x The number of rows is M y The number of columns of the modulation unit is N. x The number of rows is N y The region of modulation units on the modulation unit array detected by each detector pixel before merging is called the pre-merging detection region, and the region of equivalent modulation units on the modulation unit array detected by each detector pixel after merging is called the post-merging detection region; the side length of the pre-merging detection region is R = N. x / M x =N y / M y The dimension is R×R; the side length of the merged detection region is R. merge Dimension R merge ×R merge Determine the merging factor MG = R / R merge MG is an integer greater than 1; the number of columns of the equivalent modulation unit is N. merge，x =N x / MG, number of lines is N merge,y =N y / MG; the modulation resolution of the modulation unit array is N. merge，x ×N merge，y ;

[0015] Step 2: Set the number of snapshots L according to the required compression ratio for measurement. merge ; on the nth λ The observation matrix on the modulation unit array corresponding to the first detector pixel in each spectral channel after merging is:

[0016]

[0017] Where i is from 1 to M x ·M y integers, where l is from 1 to L merge integers, for The row vector represents the nth row. λ In the l-th frame snapshot of the spectral channels, the i-th pixel on the detector and the R on the spectral image merge ×R merge The correspondence between the merged detection regions of the size, where the superscript T indicates the transpose of the matrix; Indicates the nth λ The actual observation matrix for each spectral channel is:

[0018]

[0019] Where 0 is the zero matrix;

[0020] Step 3: Convert the row vector Stacked as R merge ×R merge A matrix of size n, obtain the nth... λ The actual coding matrix at the time of the l-th frame snapshot under each spectral channel

[0021] Step 4, according to Get the nth λ The R×R size encoding matrix loaded on the modulation unit array corresponding to the i-th detector pixel in the l-th frame snapshot under each spectral channel. according to It is possible to adjust the modulation resolution without changing the modulation unit array hardware.

[0022] Preferably, it further includes a super-resolution reconstruction module, used to obtain a higher resolution image of the target scene from the hyperspectral image of the target scene at the modulation resolution obtained by the reconstruction module through a compressed sensing reconstruction algorithm, thereby achieving super-resolution reconstruction; the super-resolution reconstruction is: solving... in, The nth modulation unit array before merging λ The vector form of the target image under each spectral channel, D dsample For downsampling matrix, The nth modulation unit array after merging λ The target image in vector form under each spectral channel, in Characterizing the partitioned in-situ modulation process; solving using a compressed sensing reconstruction algorithm:

[0023]

[0024] Where Ψ is a sparse base. Let β be the sparse coefficients corresponding to the sparse basis, and β be the noise limit; the sparse coefficient vector is obtained by solving. Depend on Recover the nth λ High spatial resolution images of the target scene across multiple spectral channels; after super-resolution reconstruction, the imaging method achieves a spatial compression ratio γ. o =(1 / R) 2 L merge .

[0025] Better place, in, It is the corresponding downsampling submatrix of the detection region corresponding to the i-th detector pixel.

[0026] Preferably, the reconstruction module reconstructs the hyperspectral image of the target scene at the modulation resolution as follows:

[0027] The nth modulation unit array after merging λ The vector form of the target image under each spectral channel. For the nth λ Measurements of the array detector under each spectral channel For the nth λ Actual observation matrix for each spectral channel To measure noise, the sampling process of the area array detector is as follows: Based on the sparsity assumption, it is expressed as the product of the sparse basis and the corresponding sparse coefficients. Among them, Ψ merge It is a sparse base. The corresponding sparsity coefficient;

[0028] Solve for the nth... λ The sparse coefficient vector corresponding to each spectral channel The inverse optimization problem is then solved to obtain the hyperspectral image of the target scene at the modulation resolution:

[0029]

[0030] Where β is the noise limit.

[0031] A variable resolution compressed sensing hyperspectral computational imaging device includes: a modulation unit array and a control module;

[0032] The control module is used to output coded signals to the modulation unit array according to the partitioned same-value modulation method;

[0033] The modulation unit array is used to modulate the light signal emitted by the target scene that converges on it according to the encoded signal and then project it onto the area array detector; each modulation unit in the modulation unit array independently receives the encoded signal output by the control module;

[0034] The partitioned same-value modulation method involves dividing the modulation unit array into sub-blocks of the same size and without overlap, based on the actual required image spatial resolution. The control module outputs the same encoded signal to the modulation units in each sub-block, so that the modulation units in the same sub-block modulate the optical signal in the same way. This allows all the modulation units in each sub-block to be merged into an equivalent modulation unit, thereby realizing the variable modulation resolution of the modulation unit array.

[0035] Beneficial effects:

[0036] 1. This invention divides the modulation units in the modulation unit array into sub-blocks of the same size and without overlap, and loads the same encoding value in each sub-block, so that the modulation units in the same sub-block modulate the optical signal in the same way. Each sub-block can be equivalently merged into a new modulation unit. This can realize the actual resolution change of the modulation unit array without changing the system hardware and structure, and thus make the spatial resolution of the image obtained by the hyperspectral computational imaging method flexible and variable.

[0037] 2. This invention, through partitioned isotropic modulation, can acquire and reconstruct spectral images of the target scene at different resolutions. Under the same compression ratio, fewer samples are required after partitioning. In addition, the dimension of the observation matrix is ​​reduced, and the reconstruction complexity is also reduced. At the same time, it can comprehensively consider the requirements of different application scenarios for target resolution, acquisition time and computational complexity, and flexibly adjust the spatial resolution of imaging by selecting appropriate merging factors.

[0038] 3. In this invention, if higher image resolution is required, an image with the same resolution as the original modulation unit array can be obtained through super-resolution reconstruction. At the same compression ratio, compared to directly obtaining a high-resolution image from the modulation unit array, the super-resolution reconstruction method in this invention has lower computational complexity, shorter reconstruction time, and improved reconstruction quality.

[0039] 4. This invention utilizes D dsample The design of a matrix, its submatrices Each line contains only MG 2With one non-zero element and all other elements being zero, the computational complexity of the compressed sensing reconstruction algorithm can be greatly reduced, making the super-resolution reconstruction method less computationally complex and shorter in reconstruction time.

[0040] 5. This invention obtains a hyperspectral image at modulation resolution by solving an inverse optimization problem based on the sparse assumption through a reconstruction module. Combined with the partitioned same-value modulation method, it realizes the variable resolution of the compressed sensing hyperspectral computational imaging method. It utilizes compressed sensing theory to reduce the pressure on the data acquisition end and lower the hardware requirements. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the DMD partitioned in-line modulation process based on an embodiment of the present invention;

[0042] Figure 2(a) shows the PSNR of the reconstructed images of each spectral channel of the target scene without merging in the simulation of the embodiment of the present invention;

[0043] Figure 2(b) shows the PSNR of the reconstructed image of each spectral channel of the target scene when the merging factor is 2 in the simulation based on the embodiment of the present invention;

[0044] Figure 2(c) shows the PSNR of the reconstructed image of each spectral channel of the target scene when the merging factor is 4 in the simulation of the embodiment of the present invention;

[0045] Figure 3(a) shows the original and reconstructed images of the target scene in different spectral channels in the simulation of the embodiment of the present invention without merging;

[0046] Figure 3(b) shows the original and reconstructed images of the target scene in different spectral channels when the merging factor is 2 in the simulation based on the embodiment of the present invention;

[0047] Figure 3(c) shows the original and reconstructed images of the target scene in different spectral channels when the merging factor is 4 in the simulation based on the embodiment of the present invention.

[0048] Figure 4(a) shows a comparison of the average PSNR of the reconstructed images using the Two-step and One-step methods in the simulation of this invention;

[0049] Figure 4(b) shows a comparison of the average reconstruction time of the Two-step and One-step methods used in the simulation of this invention;

[0050] Figure 5 This is a comparison of the reconstructed spectral images obtained using the Two-step and One-step methods in the simulation of the embodiments of the present invention. Detailed Implementation

[0051] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0052] This invention provides a variable resolution compressed sensing hyperspectral computational imaging system and device, the core of which includes: a modulation unit array, a control module, an area array detector, and a reconstruction module.

[0053] The control module outputs coded signals to the modulation unit array according to the partitioned identical modulation method. The modulation unit array modulates the light signals emitted from the target scene converged upon it using the coded signals and projects them onto the area array detector. Each modulation unit in the array independently receives the coded signal output by the control module. The partitioned identical modulation method divides the modulation unit array into non-overlapping sub-blocks of the same size according to the required image spatial resolution. The control module outputs the same coded signal to the modulation units within each sub-block, ensuring that the modulation units within the same sub-block modulate the light signal identically. This allows all modulation units in each sub-block to be merged into an equivalent modulation unit, achieving variable modulation resolution for the modulation unit array. The area array detector samples the received light signals and inputs the sampling results into the reconstruction module. The reconstruction module reconstructs a hyperspectral image of the target scene at the modulation resolution based on the sampling results.

[0054] Existing hyperspectral computational imaging methods, when reconstructing resolution using compressed sensing theory, always result in an image resolution equal to that of the modulation unit array. In other words, once the parameters of the modulation unit array are determined, the spatial resolution of the image calculated by the reconstruction algorithm is also fixed, making flexible adjustment difficult. However, the partitioned identical-value modulation method employed in this invention divides the modulation units within the modulation unit array into identical, non-overlapping sub-blocks. Each sub-block is loaded with the same encoding value, ensuring that the modulation units within the same sub-block modulate the optical signal identically. This allows the modulation units within each sub-block to be merged into new equivalent modulation units. Without altering the system hardware and structure, the actual resolution of the modulation unit array can be varied, thus enabling flexible and variable spatial resolution of the image obtained by the hyperspectral computational imaging method.

[0055] The present invention will be further described in detail below with reference to an embodiment.

[0056] In this embodiment, a digital microlens array (DMD) is selected as the modulation unit array and control module. The DMD consists of an array of micromirrors, and the tilt angle of each micromirror can be individually controlled to ±12°. By rotating the corresponding micromirror, the direction of the reflected light can be changed, putting the corresponding micromirror into an "on" or "off" state, thereby generating a binary encoding template composed of "1" and "0", realizing the modulation of the optical signal in the spatial dimension. As a microelectromechanical system with electronic input and optical output, the DMD allows developers to perform high-speed, efficient, and reliable spatial light modulation. It adopts mature semiconductor manufacturing technology, and each DMD contains up to 2 million independently controlled micromirrors, making it a mature device for realizing the modulation of optical signals in the spatial dimension.

[0057] This embodiment requires the following components: an imaging lens group, a beam splitter, a relay lens group 1, a DMD, a relay lens group 2, an area array detector, a reconstruction module, and a super-resolution reconstruction module. The light signal emitted from the target scene is projected onto the beam splitter via the imaging lens group. This beam splitter is used to split the scene information in the spectral dimension. Then, the relay lens group 1 converges the split scene information onto the DMD. The DMD uses a partitioned iso-value modulation method to perform spatial dimension encoding modulation on the received scene information. Subsequently, the relay lens group 2 projects the light intensity signal, which has undergone spectral dimension splitting and spatial dimension modulation, onto the area array detector. The area array detector samples and quantizes the light intensity signal and inputs the collected compressed observation results into the reconstruction module. The hyperspectral image of the target scene is then calculated using a compressed sensing reconstruction algorithm. In this embodiment, the DMD is used for partitioned iso-value modulation, dividing the micromirror array into identical and non-overlapping sub-regions. Iso-value encoding is loaded into each sub-region, and the micromirrors within each sub-region are merged into new micromirror units, thus changing the actual resolution of the DMD. Based on the observation matrix of the corresponding dimension after DMD modulation, the sparse basis, and the compressed observation results acquired on the detector, spectral images of the target scene at different resolutions are recovered through reconstruction algorithms, allowing for flexible adjustment of the imaging spatial resolution. Subsequently, if a higher spatial resolution is required, super-resolution reconstruction can be performed again on the lower-resolution spectral images acquired by the partitioned DMD using the compressed sensing theoretical framework, reconstructing an image with the same resolution as the original DMD.

[0058] This embodiment uses a block-based compressed sensing model for compressed sampling. Each pixel on the area array detector corresponds to a sub-block region in the DMD and target scene information. Let M be the number of detector pixel columns in the vertical direction of the area array detector. x The number of pixel rows for the detector in the horizontal direction is M. y The resolution of the array detector is M. x ×M y The DMD has N columns of modulation units in the vertical direction.x The number of horizontal modulation unit rows is N y The resolution is N x ×N y The region of modulation units on the modulation unit array detected by each detector pixel before merging is called the pre-merging detection region, and the region of equivalent modulation units on the modulation unit array detected by each detector pixel after merging is called the post-merging detection region. The horizontal resolution of the pre-merging detection region is R, the vertical resolution is R', and the dimension is R×R, where R=N. x / M x =N y / M y .

[0059] To achieve variable resolution, the steps for acquiring variable resolution hyperspectral images provided by this invention specifically include:

[0060] Step 1: Determine the merging factor MG according to the needs of different applications. MG is an integer greater than 1. After the merging factor is confirmed, the actual dimensions of the merged sub-block region and the actual resolution of the DMD are also determined. Let R be the horizontal resolution of the merged detection region. merge The vertical resolution is R merge Dimension R merge ×R merge And satisfy MG = R / R merge The number of equivalent modulation unit columns after DMD merging is N. merge，x =N x / MG, with an equivalent modulation unit row count of N merge，y =N y / MG, actual resolution is M merge，x ×N merge，y .

[0061] Step 2: Set the number of snapshots L according to the required compression ratio for measurement. merge Generate at the nth λ (n λ =1,2,...,N λ In the spectral channel i, with the i-th (i = 1, 2, ..., M) x ·M y The actual observation matrix corresponding to the sub-blocks on the low-dimensional DMD after partitioning the detector pixels. in, The superscript T indicates the transpose of the matrix. The compression ratio mentioned in this invention is one of the indicators in compressed sensing technology. In this embodiment, it represents the ratio of the vector dimension of the obtained observation value to the vector dimension of the original light signal during hyperspectral imaging. The dimension is A random observation matrix is ​​used, where each element takes the value 0 or 1, and can follow either a symmetric Bernoulli distribution or a sparse random distribution. yes The row vector represents the nth row. λ In each spectral channel, the i-th pixel on the detector in the l-th frame snapshot corresponds to the R-value on the corresponding spectral image. merge ×R merge The relationship between sub-blocks of different sizes, The number field that the quantity belongs to.

[0062] use Indicates the nth λ Actual observation matrix for each spectral channel and The relationship between them is represented by the following formula:

[0063]

[0064] in, yes zero-order matrix;

[0065] Step 3: Convert the actual observation matrix row vectors in Stacked as R merge ×R merge A matrix of size n can be used to obtain the nth matrix. λ The actual coding matrix at the time of the l-th frame snapshot under each spectral channel. Specifically, record Its composition is as follows:

[0066]

[0067] Step 4, according to Get the nth λ In each spectral channel, the R×R size encoding matrix loaded on the same sub-block region of the DMD corresponding to the i-th detector pixel at the l-th frame snapshot. Its specific composition is given by the following formula:

[0068]

[0069] Each sub-region has a size of MG×MG. All sub-matrices... By piecing them together, we obtain the nth... λ The size used in the l-th frame for each spectral channel is N. x ×N y Complete encoding template for dimensions

[0070] The partitioned in-slot modulation process of a certain sub-block in DMD is as follows: Figure 1 As shown, the merging factor is set to 2. By applying the same encoding to each 2×2 sub-region, the micromirrors within that region are rotated in the same direction. This process effectively combines the four micromirrors into a new equivalent micromirror, reducing the spatial resolution of the DMD from 8×8 to 4×4. The dimension of the corresponding encoding matrix also changes accordingly. Consequently, the observation matrix used for compressed sampling of the partitioned, in-slot modulated DMD also becomes smaller.

[0071] Step 5: During compressed sensing hyperspectral computational imaging, load the encoding template frame by frame. To achieve DMD partitioned isotropic modulation and obtain compressed measurement values ​​of hyperspectral data cubes.

[0072] Step 6: Based on the observation matrix of the corresponding dimension after DMD modulation, the sparse basis, and the compressed observation results acquired on the detector, the spectral images of the target scene at different resolutions are reconstructed using a reconstruction algorithm.

[0073] use Indicates the nth partition of the DMD λ Target scene information for each spectral channel The measured value representing this spectral channel, the process of compressed sampling is expressed as: in, This represents measurement noise. Based on the sparsity assumption, It can be expressed as the product of the sparse basis and the corresponding sparse coefficients: in, It is a sparse base. This represents the sparsity coefficient.

[0074] By solving the following inverse optimization problem, the nth digit can be reconstructed. λ The sparse coefficient vectors corresponding to each spectral channel are used to reconstruct the spectral image for that channel.

[0075]

[0076] Where β represents the noise limit. After partitioned isotropic modulation, the resolution of DMD is increased from N. x ×N y Change to N merge,x ×N merge,y Therefore, the spatial resolution of the target scene it collects is also increased from N. x ×N y Change to N merge，x ×N merge，y At this point, the overall compression ratio in the spatial dimension is γ. o =(1 / R) merge ) 2 Lmerge It is evident that, at the same compression ratio, the partitioned low-resolution DMD requires fewer samples, and with the reduced dimension of the observation matrix, the reconstruction complexity decreases accordingly. Researchers can flexibly acquire hyperspectral images of different resolutions by comprehensively considering the different requirements of target resolution, acquisition time, and computational complexity for different application scenarios and selecting appropriate merging factors.

[0077] After completing step 6, if a higher demand for image spatial resolution arises, the present invention can further perform super-resolution reconstruction step 7:

[0078] Step 7: By solving the following underdetermined problem, perform super-resolution reconstruction again on the lower-resolution spectral image obtained from the partitioned DMD:

[0079]

[0080] in, When DMD is not merged, the nth λ The vector form of the target image for each spectral channel. Represents the downsampling matrix. The partitioned in-situ modulation process of DMD was characterized. Furthermore, D... dsample The specific components are as follows:

[0081]

[0082] in, yes zero-order matrix This represents the downsampling sub-matrix applied to the target scene sub-block corresponding to the i-th detector pixel. It is generally assumed that the downsampling operation is linear, and its specific form is given below:

[0083]

[0084] in, It is given by the following formula:

[0085]

[0086] Since it is highly uncorrelated with most sparse bases, super-resolution reconstruction can be achieved by solving the underdetermined problem using compressed sensing theory. Compressed sensing reconstruction algorithms, including GPSR or TwIST, are used to solve the inverse optimization problem.

[0087]

[0088] in, β represents a sparse basis, and β is the noise limit. Let represent the corresponding sparse coefficients, and β be the noise limit. Solving for the sparse coefficient vector yields the sparse representation relation. Recover the nth λ Spectral images under each spectral channel.

[0089] D dsample submatrix Each line contains only MG 2 With one non-zero element and all other elements being zero, the computational cost of compressed sensing reconstruction algorithms can be significantly reduced. After super-resolution reconstruction is completed, the compression ratio of the method described in this application in the spatial dimension becomes γ. o =(1 / R) 2 L merge Under the same compression ratio, compared with directly acquiring high-resolution images from the original DMD, the super-resolution reconstruction method proposed in this invention has lower computational complexity, shorter reconstruction time, and improved reconstruction quality.

[0090] The following simulation demonstrates the effectiveness and superiority of the variable resolution compressed sensing hyperspectral computational imaging method.

[0091] The programming platform used for the simulation experiment was MATLAB R2016a. The hyperspectral data used in the simulation experiment used Lego figures as the target objects, encompassing 24 spectral bands from 452nm to 667nm, with a spatial size of 256×256 pixels. The simulated DMD contained 256×256 pixels, while the simulated detector contained 32×32 pixels. Each 8×8 pixel in the coded aperture corresponded to one detector pixel, i.e., R=8. In the simulation, Gaussian white noise with a signal-to-noise ratio (SNR) of 30dB was used to simulate detector noise. A 2D-IDCT (Two-dimensional Inverse DCT) basis was used to sparsely represent the spectral image, and the GPSR algorithm was used for reconstruction. To objectively evaluate the experimental results, the peak signal-to-noise ratio (PSNR) was used as a measure of the spectral image reconstruction quality.

[0092] The overall compression ratio γ o The merging factor was set to 0.5. When the merging factor was 2 and 4, the compressed observation values ​​of the target scene were collected through simulation steps 1-6. Then, the reconstruction algorithm was used to recover the spectral images of different resolutions from the compressed measurement values, realizing the acquisition of variable resolution hyperspectral images, and comparing them with the imaging method without micromirror unit merging.

[0093] Figures 2(a)-2(c)PSNR curves for the reconstructed images of each spectral channel of the target scene under different merging factors were plotted, corresponding to the three cases of no merging, merging factor of 2, and merging factor of 4, respectively. The solid line represents the average PSNR of five trials, while the shaded area indicates the fluctuation range of PSNR. The average PSNR of all spectral channels in the three cases were 31.79 dB, 28.83 dB, and 25.38 dB, respectively, and the reconstruction results all showed good robustness.

[0094] To visually demonstrate the reconstruction effect of the variable resolution compressed sensing hyperspectral computational imaging method, several spectral channels under different merging factors are selected as representatives, and the corresponding original and reconstructed images for each channel are given, such as... Figures 3(a)-3(c) As shown, steps 1-6 effectively achieve variable spatial resolution. It should be noted that the required number of snapshot frames decreases proportionally for the three cases: no merging, a merging factor of 2, and a merging factor of 4. This means that at the same compression ratio, the larger the merging factor, the shorter the observation time almost proportionally. Furthermore, the reconstruction complexity decreases as the merging factor increases. Without merging, the total reconstruction time is 74.81s; with a merging factor of 2, the total reconstruction time is 15.56s; and with a merging factor of 4, the total reconstruction time is 2.49s. In different application scenarios, an appropriate merging factor can be selected according to specific requirements.

[0095] Next, another set of simulations illustrates the effectiveness of step 7. This set of simulations completed target scene reconstruction for two paths: First, without merging, observations were performed using an 8-frame coded template, and a spectral image with a spatial resolution of 256×256 was reconstructed from the compressed observations; this is denoted as the One-step method. Second, with a merging factor of 2, observations were first performed using an 8-frame coded template, and a spectral image with a spatial resolution of 128×128 was reconstructed from the compressed observations. Then, a compressed sensing reconstruction algorithm was used to solve the underdetermined problem, and a spectral image with a spatial resolution of 256×256 was reconstructed from the 128×128 image; this is denoted as the Two-step method. The overall compression ratio for both methods is γ. o =0.125.

[0096] The average PSNR and average time for each spectral channel of the data cubes acquired by the two methods are shown in Figures 4(a) and 4(b), respectively. The average PSNR of the hyperspectral data cube acquired using the One-step method is 25.54 dB, with a total reconstruction time of 46.75 s; the average PSNR of the hyperspectral data cube acquired using the Two-step method is 27.57 dB, with a total reconstruction time of 26.43 s. It is important to note that the total time for both the One-step and Two-step methods consists of observation time and reconstruction time. The observation times are essentially the same for both methods; the difference lies in the reconstruction time. The reconstruction time for the Two-step method includes the time for reconstructing the lower-resolution image by compressing the observations and the time for super-resolution reconstruction.

[0097] Figure 5 Simulation results for two methods are presented for the Lego data cube in four spectral channels with center wavelengths of 459 nm, 511 nm, 563 nm, and 615 nm. It is evident that, at the same compression ratio, the Two-step method achieves better reconstruction quality than the One-step method. Figure 5 The demonstration also showcased magnified details of the area around Lego's eyes, highlighting the significant improvement in reconstruction quality achieved by the Two-step method. Simulation results show that the Two-step method has a shorter computation time and achieves better reconstruction results compared to the One-step method.

[0098] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A variable resolution compressed sensing hyperspectral computational imaging system, characterized in that, include: Modulation unit array, control module, area array detector, and reconstruction module; The control module is used to output coded signals to the modulation unit array according to the partitioned same-value modulation method; The modulation unit array is used to modulate the light signal emitted by the target scene that converges on it according to the encoded signal and then project it onto the area array detector; each modulation unit in the modulation unit array independently receives the encoded signal output by the control module; The array detector is used to sample the received optical signal and input the sampling result into the reconstruction module; The reconstruction module is used to reconstruct a hyperspectral image of the target scene at modulation resolution based on the sampling results; The partitioned in-situ modulation method includes the following steps: Step 1: The number of detector pixel columns of the area array detector is M. x The number of rows is M y The number of columns of the modulation unit is N. x The number of rows is N y The region of modulation units on the modulation unit array detected by each detector pixel before merging is called the pre-merging detection region, and the region of equivalent modulation units on the modulation unit array detected by each detector pixel after merging is called the post-merging detection region. The side length of the detection region before merging is R = N. x / M x =N y / M y The dimension is R×R; the side length of the merged detection region is R. merge Dimension R merge ×R merge Determine the merging factor MG = R / R merge MG is an integer greater than 1; the number of columns of the equivalent modulation unit is N. merge，x =N x / MG, number of lines is N merge，y =N y / MG; the modulation resolution of the modulation unit array is N. merge，x ×N merge，y ; Step 2: Set the number of snapshots L according to the required compression ratio for measurement. merge ; on the nth λ The observation matrix on the modulation unit array corresponding to the i-th detector pixel in each spectral channel after merging is: Where i is from 1 to M x ·M y integers, where l is from 1 to L merge integers, for The row vector represents the nth row. λ In the l-th frame snapshot of the spectral channels, the i-th pixel on the detector and the R on the spectral image merge ×R merge The correspondence between the merged detection regions of the size, where the superscript T indicates the transpose of the matrix; Indicates the nth λ The actual observation matrix for each spectral channel is: Where 0 is the zero matrix; Step 3: Convert the row vector Stacked as R merge ×R merge A matrix of size n, obtain the nth... λ The actual coding matrix at the time of the l-th frame snapshot under each spectral channel Step 4, according to Get the nth λ The R×R size encoding matrix loaded on the modulation unit array corresponding to the i-th detector pixel at the l-th frame snapshot under each spectral channel. according to It is possible to adjust the modulation resolution without changing the modulation unit array hardware.

2. The variable resolution compressed sensing hyperspectral computational imaging system as described in claim 1, characterized in that, It further includes a super-resolution reconstruction module, which is used to obtain a higher resolution image of the target scene from the hyperspectral image of the target scene at the modulation resolution obtained by the reconstruction module, and to achieve super-resolution reconstruction.

3. The variable resolution compressed sensing hyperspectral computational imaging system as described in claim 2, characterized in that, The super-resolution reconstruction is: solving... in, The nth modulation unit array before merging λ The vector form of the target image under each spectral channel, D dsample For downsampling matrix, The nth modulation unit array after merging λ The target image in vector form under each spectral channel, in Characterizing the partitioned in-situ modulation process; solving using a compressed sensing reconstruction algorithm: Where Ψ is a sparse base. Let β be the sparse coefficients corresponding to the sparse basis, and β be the noise limit; the sparse coefficient vector is obtained by solving. Depend on Recover the nth λ High spatial resolution images of the target scene across multiple spectral channels; after super-resolution reconstruction, the imaging method achieves a spatial compression ratio γ. o =(1 / R) 2 L merge .

4. The variable resolution compressed sensing hyperspectral computational imaging system as described in claim 3, characterized in that, in, It is the corresponding downsampling submatrix of the detection region corresponding to the i-th detector pixel.

5. The variable resolution compressed sensing hyperspectral computational imaging system as described in any one of claims 1-4, characterized in that, The reconstruction module reconstructs the hyperspectral image of the target scene at the modulation resolution as follows: The sampling process of the area array detector is as follows: in, The nth modulation unit array after merging λ The vector form of the target image under each spectral channel. For the nth λ Measurements of the array detector under each spectral channel For the nth λ Actual observation matrix for each spectral channel To measure noise, based on the sparsity assumption, it is represented by the product of the sparse basis and the corresponding sparse coefficients. Among them, Ψ merge It is a sparse base. The corresponding sparsity coefficient; Solve for the nth... λ The sparse coefficient vector corresponding to each spectral channel The inverse optimization problem is then solved to obtain the hyperspectral image of the target scene at the modulation resolution: Where β is the noise limit.

Images