Method and apparatus for robustness evaluation of depth reconstruction model of hyperspectral computed imaging

Abstract

The present application relates to a hyperspectral computed imaging depth reconstruction model robustness evaluation method and device, and relates to the technical field of model evaluation. The method comprises: obtaining an original hyperspectral data cube to be evaluated and a target physical mask; based on a momentum iterative optimization framework, an enhanced mask strategy is integrated according to an initial test disturbance and an initial momentum term, and an optimized test disturbance is generated through multiple iterations; the optimized test disturbance is applied to the original hyperspectral data cube to generate a final test sample; the final test sample is input into the depth reconstruction model to be evaluated to generate a reconstruction result, and the robustness of the model is quantitatively evaluated by comparing the evaluation index difference between the reconstruction result and the original hyperspectral data cube. The present application is specially aimed at enhancing the diversity of the mask, and can more comprehensively and rigorously test the robustness of the model.

Description

Technical Field

[0001] This invention relates to the field of model evaluation technology, and in particular to a method and apparatus for evaluating the robustness of depth reconstruction models in hyperspectral computational imaging. Background Technology

[0002] Hyperspectral computational imaging is a technique that uses physical encoding to compress three-dimensional hyperspectral scene data into two-dimensional measurements, and then reconstructs the original information through computation. Its core comprises a forward encoding process implemented using physical masks (such as coded apertures and optical filter arrays), and an inverse problem-solving model based on compressed sensing. The physical mask modulates the input hyperspectral data cube with a specific spatial-spectral mode to generate a two-dimensional compressed measurement, while the reconstruction model is responsible for reconstructing the original three-dimensional hyperspectral image from this compressed measurement with high quality. Deep learning-based reconstruction models, with their powerful mapping learning capabilities, have become a key technology for achieving high-quality reconstruction.

[0003] However, while deep learning models possess excellent reconstruction performance, their decision boundaries are complex and susceptible to small, carefully designed input perturbations. Adversarial examples generated by such perturbations can expose the model's potential vulnerabilities. Therefore, evaluating model robustness using adversarial examples has become a core method for measuring its reliability in complex real-world environments. In fields such as image recognition, classic adversarial example generation methods such as Projected Gradient Descent (PGD), Momentum Iterative Fast Gradient Signed Method (MI-FGSM), and Diverse Input Method (DIM) have been developed to reveal the model's decision boundaries. However, when these general methods are directly applied to deep reconstruction models for hyperspectral computational imaging, they exhibit significant adaptability limitations. This is because such models embed unique physical imaging pathways, which are fundamentally different from the application scenarios of ordinary image recognition models.

[0004] Existing evaluation methods typically add perturbations directly to the compressed 2D measurements in the digital domain. However, real-world interference sources (such as changes in illumination and perturbations to the reflectivity of object surfaces) often originate from the hyperspectral scene itself. This fails to construct a complete evaluation chain from the 3D hyperspectral scene perturbation through physical mask forward encoding to obtain 2D measurements, and finally to the degradation of the reconstructed model's performance. This results in a disconnect between the evaluation scenario and real physical threats, limiting the practical guidance value of the evaluation conclusions. Furthermore, the core of a hyperspectral computational imaging system is the physical encoding mask, which determines the information compression method. Comprehensive robustness evaluation requires verifying the model's robustness in unknown systems (unknown masks, unknown model parameters). This necessitates that evaluation methods generate highly transferable adversarial examples. However, existing methods lack effective mechanisms to improve this transferability, particularly failing to explicitly model and utilize the distribution characteristics of the mask to enhance the generalization ability of test samples. Consequently, they cannot support rigorous, black-box-like cross-system robustness comparisons and benchmark tests. Summary of the Invention

[0005] Therefore, it is necessary to provide a method and apparatus for evaluating the robustness of a depth reconstruction model for hyperspectral computational imaging that can specifically enhance mask diversity during the process of fusion physical imaging, in order to address the aforementioned technical problems.

[0006] A robustness evaluation method for depth reconstruction models in hyperspectral computational imaging, the method comprising:

[0007] Step 1: Obtain the original hyperspectral data cube and target physical mask to be evaluated;

[0008] Step 2: Based on the initial test perturbation and initial momentum term, the optimized test perturbation is generated through multiple iterations by incorporating a mask enhancement strategy based on the momentum iterative optimization framework.

[0009] Step 3: Apply the optimized test perturbation to the original hyperspectral data cube to generate the final test sample;

[0010] Step 4: Input the final test sample into the depth reconstruction model to be evaluated to generate reconstruction results. By comparing the differences in evaluation indicators between the reconstruction results and the original hyperspectral data cube, the robustness of the model is quantified.

[0011] On the other hand, a robustness evaluation device for a depth reconstruction model of hyperspectral computational imaging is also provided, comprising:

[0012] The data acquisition module is used to acquire the raw hyperspectral data cube and target physical mask to be evaluated;

[0013] The iterative optimization module is used to generate an optimized test perturbation based on the initial test perturbation and initial momentum term, using a momentum iterative optimization framework and incorporating a mask enhancement strategy, through multiple iterations.

[0014] The test sample generation module is used to apply the optimized test perturbation to the original hyperspectral data cube to generate the final test sample.

[0015] The evaluation module is used to input the final test sample into the depth reconstruction model to be evaluated, generate reconstruction results, and quantify the robustness of the evaluation model by comparing the differences between the reconstruction results and the evaluation indicators of the original hyperspectral data cube.

[0016] Compared with existing technologies, the method and apparatus for robustness evaluation of depth reconstruction models in hyperspectral computational imaging provided by this invention have the following advantages:

[0017] 1. The test perturbation is applied to the hyperspectral scene data cube before the physical imaging process. The model performance degradation is evaluated through a complete process of "perturbation scene, physical coding, compressed measurement, and model reconstruction". This link is consistent with the location of the interference source in the actual physical world, making the evaluation results more practically instructive.

[0018] 2. By enhancing mask diversity through mask augmentation strategies, a large number of diverse physical mask encoding patterns can be dynamically simulated during test perturbation generation. This enables the generated test samples to learn to adapt to changes in mask distribution, greatly improving their transferability and generalization ability across different masks and unknown imaging systems. Robustness evaluation based on these samples can more comprehensively and rigorously examine the model's true robustness in open and variable physical environments.

[0019] 3. The mask enhancement strategy and overall evaluation framework proposed in this invention do not depend on specific depth reconstruction model structures or physical imaging system implementation details, and can be easily integrated into existing hyperspectral computational imaging robustness evaluation processes. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention, and those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0021] Figure 1 A schematic flowchart illustrating the robustness evaluation method for the depth reconstruction model of hyperspectral computational imaging provided in Example 1;

[0022] Figure 2 This is a structural block diagram of the robustness evaluation device for the depth reconstruction model of hyperspectral computational imaging provided in Example 2;

[0023] Figure 3 This is an internal structural diagram of the computer device provided in Example 3.

[0024] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0026] It should be noted that in this invention, the use of terms such as "first," "second," etc., is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0027] It is understood that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0028] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] Example 1

[0030] like Figure 1 The method for robustness evaluation of depth reconstruction model in hyperspectral computational imaging provided in this embodiment includes the following steps:

[0031] Step 1: Obtain the original hyperspectral data cube and target physical mask to be evaluated.

[0032] Step 2: Based on the initial test perturbation and initial momentum term, the optimized test perturbation is generated through multiple iterations by incorporating a mask enhancement strategy based on the momentum iterative optimization framework.

[0033] Step 3: Apply the optimized test perturbation to the original hyperspectral data cube to generate the final test sample.

[0034] Step 4: Input the final test sample into the depth reconstruction model to be evaluated to generate reconstruction results. By comparing the differences in evaluation metrics between the reconstruction results and the original hyperspectral data cube, the robustness of the model is quantified.

[0035] In the specific implementation of step 1, the original hyperspectral data cube to be evaluated is denoted as... ,in , These are the space height and width, respectively. The number of spectral channels can be obtained from publicly available hyperspectral datasets (such as the KAIST dataset) or from real-world data collection.

[0036] Target physical mask It is the core coding component of the target hyperspectral computational imaging system, corresponding to physical structures such as the coding aperture and optical filter array. Additionally, it is necessary to determine the physical forward model of the target hyperspectral computational imaging system. This model is used to describe the encoding process of hyperspectral data after physical mask modulation.

[0037] In this step, the acquired raw data must match the target physical mask to ensure the consistency and effectiveness of subsequent evaluation processes.

[0038] In the specific implementation of step 2, the initialization operation is first performed: the three-dimensional perturbation tensor is initialized to all zeros. As the initial test perturbation, initialize the three-dimensional weight gradient accumulation tensor with all zeros. As the initial momentum term; simultaneously, hyperparameters are set, including the maximum amplitude constraint of the perturbation. Iteration step size Total number of iterations Momentum decay factor and mask enhancement probability .

[0039] Then, using the initial test perturbation and the initial momentum term as initial inputs, the iterative process is executed according to a preset total number of iterations. The input for each iteration is the test perturbation output from the previous iteration and the updated momentum term.

[0040] In the During this iteration, the following sub-steps are executed:

[0041] Step 201: Sample the target physical mask and the preset mask enhancement distribution according to the mask enhancement probability. The physical mask used in the next iteration.

[0042] Specifically, based on the mask enhancement probability of the current iteration, the mask sampling function is called to sample the physical mask used in the current iteration from the target physical mask and the preset mask enhancement distribution. The process expression is as follows:

[0043] ;

[0044] In the formula, Indicates the first The physical mask used in the next iteration; Indicates the mask sampling function; Represents the target physical mask; This indicates a pre-defined mask enhancement distribution; Indicates the probability of mask enhancement; This represents the mask sampled from the mask enhancement distribution.

[0045] Step 202, according to the first The physical mask used in the next iteration Raw hyperspectral data cube With the Test perturbation after the second iteration optimization Through forward imaging and reconstruction, a reconstructed hyperspectral data cube is output.

[0046] Specifically, at the beginning of the iteration, let the initial test perturbation be set. Set it to 0; then, set the first... Test perturbation after the second iteration optimization Add to the original hyperspectral data cube Obtain hyperspectral data containing perturbations. .

[0047] Then, through physical masking For physical imaging systems The system is then parameterized to obtain the parameterized physical imaging system.

[0048] Perturbed hyperspectral data obtained through a parameterized physical imaging system Encode and output analog compressed measurement .

[0049] Simulated compression measurement Input the depth reconstruction model to be evaluated Process the data to output a reconstructed hyperspectral data cube. .

[0050] Step 203: Calculate the first... The loss of the nth iteration, and according to the nth iteration loss. The corresponding iteration loss calculation for the [number]th iteration The gradient of the next iteration.

[0051] Specifically, at the beginning of the iteration, let the initial test perturbation be set. The value is 0; then the reconstructed hyperspectral data cube is calculated. With perturbation-containing hyperspectral data The difference between them is the first Loss of the next iteration Then calculate the loss. Compared to perturbated hyperspectral data iterative gradient .

[0052] Step 204, according to the preset momentum decay factor, the first The momentum term after the second iteration update is the same as the first iteration. The gradient of the nth iteration is updated through momentum accumulation to obtain the th... The momentum term after the next iteration update.

[0053] Specifically, at the beginning of the iteration, let the initial momentum be... The value is 0; then the momentum decay factor is used. , No. The momentum term after the second iteration update is the same as the first iteration. The gradient of the nth iteration is obtained by updating the momentum accumulation. The momentum term updated in the next iteration is expressed as follows:

[0054] ;

[0055] In the formula, Indicates the first The momentum term updated in the next iteration; Indicates the momentum decay factor; Indicates the first The momentum term updated in the next iteration; Indicates the first The gradient of the next iteration; Indicates the first Loss per iteration; express Norms are used to normalize gradients.

[0056] Step 205, according to the first The momentum term updated in the next iteration, the first The test perturbation after the second iteration optimization is adjusted along the momentum term direction and clipped to the constraint range, with the preset iteration step size and maximum perturbation amplitude as constraints, and the result is the output of the second iteration. The test perturbation after the next iteration optimization.

[0057] Specifically, with a preset iteration step size and maximum perturbation amplitude as constraints, the perturbation is adjusted along the direction of the current momentum term. The gradient direction is determined by the sign function, and then the perturbation is clipped to the constraint range by the clipping function, outputting the first iteration step. The test perturbation after the next iteration of optimization is expressed as follows:

[0058] ;

[0059] In the formula, Indicates the first The test perturbation after the next iteration optimization; Represents the clipping function; Indicates the first The test perturbation after the next iteration optimization; Indicates the preset iteration step size; Represents a symbolic function; Represents a cube containing the raw hyperspectral data; Indicates the maximum amplitude of the disturbance; Indicates the first The momentum term after the next iteration update.

[0060] Step 206, with the first The test perturbation after the second iteration optimization and the first The momentum term updated in the next iteration is used as input. Steps 201 to 205 are repeated until the preset total number of iterations is reached, and the optimized test perturbation is output.

[0061] It is understandable that the first Test perturbation after the second iteration optimization With the The momentum term after the next iteration update As input for the next iteration, repeat steps 201 to 205 until the preset total number of iterations N is reached. At this point, the output... This refers to the optimized test perturbation.

[0062] Furthermore, mask enhancement distribution Based on the type definition of the target hyperspectral computational imaging system. If the target hyperspectral computational imaging system is a coded aperture snapshot spectral imaging system, the mask enhancement distribution... For binary mask distribution With grayscale mask distribution A uniform mixture is represented as .

[0063] Among them, the binary mask distribution The generation process is as follows: First, in the interval Uniformly sample a transmittance parameter Then based on the sampled transmittance parameters Generate a size of Basic binary mask matrix In the basic binary mask matrix In the example, each element independently follows the parameter... The Bernoulli distribution; then the basic binary mask matrix Upsampling to target spatial size This yields the final binary mask distribution. In the formula, This represents the minimum transmittance. This indicates the maximum transmittance.

[0064] Grayscale mask distribution The generation process is as follows: First, in the interval and Uniformly sampled mean values ​​within each cell and standard deviation Then based on the mean and standard deviation Generate a size of Basic grayscale mask matrix In the basic grayscale mask matrix In the expression, each element independently follows the mean. Standard deviation is The Gaussian distribution; then, the basic grayscale mask matrix... The cropping process is performed, and the cropping range is set according to requirements. In this embodiment, it is set to... Finally, the cropped base grayscale mask matrix is ​​upsampled to the target space size. This yields the final grayscale mask distribution. In the formula, This represents the minimum value of the mean; This represents the maximum value of the mean. This represents the minimum standard deviation. This represents the maximum standard deviation. It's worth noting that in grayscale mask distribution... In the mean It is considered as a transmittance parameter.

[0065] If the target hyperspectral computational imaging system is an imaging system using optical filters, the mask enhancement distribution... Three-dimensional grayscale mask distribution The generation process is as follows: First, in the interval... and Uniformly sampled mean values ​​within each cell and standard deviation Then based on the mean and standard deviation Generate a size of Basic 3D mask tensor In the basic three-dimensional mask tensor In the expression, each element independently follows the mean. Standard deviation is The Gaussian distribution; then, the basic 3D mask tensor The cropping process is performed, and the cropping range is set according to requirements. In this embodiment, it is set to... Finally, the cropped base 3D mask tensor is upsampled to the target space and spectral dimensions. This yields the final three-dimensional grayscale mask distribution.

[0066] It is worth noting that the binary mask distribution Grayscale mask distribution With 3D grayscale mask distribution The base mask size in the middle satisfies , , scaling factor This is used to simulate the spatial correlation of encoded values ​​between adjacent pixels caused by the physical mask unit size being larger than the sensor pixel size.

[0067] In the specific implementation of step 3, the optimized test perturbation will be... Apply additively to the original hyperspectral data cube Generate the final test samples for robustness evaluation. .

[0068] For the final test samples Furthermore, the physical forward model of the target hyperspectral imaging system can be calculated. Encode the data to generate the corresponding evaluation compression measurement. The evaluation compression measurement It can be directly used as input to the deep reconstruction model to be evaluated, adapting to the actual input format requirements of the model.

[0069] In the specific implementation of step 4, the final test sample generated in step 3 will be... Or its corresponding evaluation compression measurement Input the depth reconstruction model to be evaluated The reconstruction results were obtained.

[0070] The evaluation metric used is the average rate of decline in PSNR. Specifically, the decline rate is first calculated for all test samples. The PSNR decline rate (PDR) is calculated, and then the arithmetic mean of all PDRs is taken to obtain the average PSNR decline rate. A higher value indicates a more significant decrease in model performance caused by the test samples. The weaker the robustness, the better. Quantitative indicators can intuitively reflect the robustness of the model when facing test samples with masked perturbations, enabling an objective and accurate assessment of the model's robustness.

[0071] Among them, test samples The PDR calculation expression is:

[0072] ;

[0073] In the formula, This represents the PSNR value of the original hyperspectral data cube and its reconstruction result; This represents the PSNR value between the test sample and its reconstruction result; This represents the reconstruction result of the original hyperspectral data; This indicates the reconstruction result of the test sample.

[0074] In one embodiment, specific experimental results of robustness evaluation using the method proposed in this invention are presented and compared with existing methods to verify the effectiveness of the method proposed in this invention.

[0075] The experiment generated test samples on 10 outdoor hyperspectral scenes from the KAIST dataset. To verify the cross-model and cross-mask transferability of the test samples, various mask and model configurations were used. Regarding mask configuration, in addition to the original target mask (RAW) used to generate the test samples, several new mask configurations were introduced for transfer evaluation to simulate the uncertainties of real-world systems. For example, p0.5-s128- represents a randomly generated binary mask with a transmittance of 0.5 and a base size of 128×128. The evaluation models included MST-L, TSA-Net, and PADUT models. Five typical methods were selected for comparison: MI-FGSM, CW, two classic input augmentation methods, DIM and MIG, and the method proposed in this invention. The evaluation metric was the average PSNR decrease rate. The PDR of the 10 test samples was calculated using the PDR calculation expression, and then the arithmetic mean was taken to obtain the final average PSNR decrease rate.

[0076] Table 1 shows the evaluation results of test samples generated on MST-L under different mask and model configurations in the CASSI system. The first row represents the white-box results of the test samples, indicating that the test samples were generated under this configuration. In the white-box evaluation scenario, samples generated by the traditional CW method cause the greatest performance degradation, at 67.81%. At this time, the average PSNR decrease rate of the proposed method is 57.29%, which is lower than that of the traditional method optimized for fixed configurations, but still has significant evaluation capability. This is consistent with the design intention of the mask enhancement strategy to improve generalization rather than pursue the ultimate performance of a single configuration.

[0077] In cross-model and cross-mask migration evaluation scenarios that better reflect real-world black-box conditions, the advantages of the proposed method are fully demonstrated. When the target model is MST-L and the mask becomes an unknown configuration p0.5-s128-grayscale, the best-performing traditional method, MIG, has an average PSNR decrease rate of 42.66%, while the proposed method achieves an average PSNR decrease rate as high as 54.42%, significantly outperforming all comparative methods.

[0078] Under various combinations of transmittance, base size, and mask type (binary or grayscale), the method proposed in this invention consistently achieved the highest average PSNR decrease rate, demonstrating that the generated test samples have excellent mobility and strong adaptability to mask diversity.

[0079] Table 1. Average PSNR decrease rate (%) for test samples under different mask and model configurations

[0080]

[0081] Experimental results demonstrate that the proposed method generates test samples with stronger transferability and more significant evaluation effects when assessing the robustness of hyperspectral computational imaging models to unknown mask configurations. Compared to traditional methods that directly augment in the data domain, the strategy of augmenting in the physical mask space more effectively simulates the uncertainties of real systems, thus enabling the evaluation conclusions to more comprehensively and rigorously reflect the true robustness of the model under complex and variable physical environments. This verifies the effectiveness and advancement of this invention in improving the generalization and practical guidance value of robustness assessment.

[0082] It should be understood that, although this embodiment Figure 1 The steps are shown sequentially as indicated by the arrows, but they are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are performed; they can be executed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0083] Example 2

[0084] Based on the robustness evaluation method for depth reconstruction models in hyperspectral computational imaging in Example 1, this example discloses a robustness evaluation device for depth reconstruction models in hyperspectral computational imaging, such as... Figure 2 As shown, the robustness evaluation device for the depth reconstruction model of hyperspectral computational imaging includes: a data acquisition module 501, an iterative optimization module 502, a test sample generation module 503, and an evaluation module 504, wherein:

[0085] The data acquisition module 501 is used to acquire the original hyperspectral data cube and target physical mask to be evaluated.

[0086] The iterative optimization module 502 is used to generate an optimized test perturbation based on the initial test perturbation and initial momentum term, using a momentum iterative optimization framework and incorporating a mask enhancement strategy, through multiple iterations.

[0087] The test sample generation module 503 is used to apply the optimized test perturbation to the original hyperspectral data cube to generate the final test sample.

[0088] Evaluation module 504 is used to input the final test sample into the depth reconstruction model to be evaluated, generate reconstruction results, and quantify the robustness of the model by comparing the evaluation index difference between the reconstruction results and the original hyperspectral data cube.

[0089] In this embodiment, the specific working process and working principle of the data acquisition module 501, iterative optimization module 502, test sample generation module 503, and evaluation module 504 are the same as those in Embodiment 1, and therefore will not be described again in this embodiment. Each unit module can be implemented entirely or partially through software, hardware, or a combination thereof. Each unit module can be embedded in or independent of the processor in the computer device in hardware form, or it can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above unit modules.

[0090] Example 3

[0091] like Figure 3 The diagram illustrates a terminal device disclosed in this embodiment, comprising a transmitter, a receiver, a memory, and a processor. The transmitter transmits instructions and data, the receiver receives instructions and data, the memory stores computer-executed instructions, and the processor executes the computer-executed instructions stored in the memory to implement the method described in Embodiment 1 above.

[0092] It is important to note that the aforementioned memory can be either standalone or integrated with the processor. When the memory is set up independently, the terminal device also includes a bus for connecting the memory and the processor.

[0093] Example 4

[0094] This embodiment discloses a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the method in Embodiment 1 above.

[0095] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0096] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0097] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for evaluating the robustness of a depth reconstruction model in hyperspectral computational imaging, characterized in that, The method includes: Step 1: Obtain the original hyperspectral data cube and target physical mask to be evaluated; Step 2: Based on the initial test perturbation and initial momentum term, the optimized test perturbation is generated through multiple iterations by incorporating a mask enhancement strategy based on the momentum iterative optimization framework. Step 3: Apply the optimized test perturbation to the original hyperspectral data cube to generate the final test sample; Step 4: Input the final test sample into the depth reconstruction model to be evaluated to generate reconstruction results. By comparing the differences in evaluation metrics between the reconstruction results and the original hyperspectral data cube, the robustness of the model is quantified. In step 2, based on the initial test perturbation and initial momentum term, a masking enhancement strategy is incorporated, and an optimized test perturbation is generated through multiple iterations, including: Using the initial test perturbation and the initial momentum term as initial inputs, the iterative process is executed according to a preset total number of iterations. The input for each iteration is the test perturbation output from the previous iteration and the updated momentum term. The next iteration includes: Step 201: Sample the target physical mask and the preset mask enhancement distribution according to the mask enhancement probability. The physical mask used in the next iteration; Step 202, according to the first The physical mask used in the next iteration, the original hyperspectral data cube, and the... The test perturbation optimized in the next iteration is used to output a reconstructed hyperspectral data cube through forward imaging and reconstruction. Step 203: Calculate the first... based on the reconstructed hyperspectral data cube and the perturbed hyperspectral data... The loss of the nth iteration, and according to the nth iteration loss. The corresponding iteration loss calculation for the [number]th iteration The gradient of the next iteration; Step 204, according to the preset momentum decay factor, the first The momentum term after the second iteration update is the same as the first iteration. The gradient of the nth iteration is updated through momentum accumulation to obtain the th... The momentum term updated in the next iteration; Step 205, according to the first The momentum term updated in the next iteration, the first The test perturbation after the second iteration optimization is adjusted along the momentum term direction and clipped to the constraint range, with the preset iteration step size and maximum perturbation amplitude as constraints, and the result is the output of the second iteration. The test perturbation after the next iteration optimization; Step 206, with the first The test perturbation after the second iteration optimization and the first The momentum term updated in the next iteration is used as input. Steps 201 to 205 are repeated until the preset total number of iterations is reached, and the optimized test perturbation is output.

2. The robustness evaluation method for depth reconstruction models in hyperspectral computational imaging according to claim 1, characterized in that, In step 201, the target physical mask and the preset mask enhancement distribution are sampled according to the mask enhancement probability. The physical mask used in this iteration is expressed as follows: ； In the formula, Indicates the first The physical mask used in the next iteration; Indicates the mask sampling function; Represents the target physical mask; This indicates a pre-defined mask enhancement distribution; Indicates the probability of mask enhancement; This represents the mask sampled from the mask enhancement distribution.

3. The robustness evaluation method for depth reconstruction models in hyperspectral computational imaging according to claim 1, characterized in that, In step 202, according to the first The physical mask used in the next iteration, the original hyperspectral data cube, and the... The test perturbation optimized in the next iteration, through forward imaging and reconstruction, outputs a reconstructed hyperspectral data cube, including: The first The test perturbation optimized in the next iteration is added to the original hyperspectral data cube to obtain hyperspectral data with perturbation; The physical imaging system is parameterized using the physical mask to obtain the parameterized physical imaging system. The perturbated hyperspectral data is encoded using the parameterized physical imaging system, and a simulated compressed measurement is output. The simulated compression measurement is input into the depth reconstruction model to be evaluated and processed to output a reconstructed hyperspectral data cube.

4. The robustness evaluation method for depth reconstruction models in hyperspectral computational imaging according to claim 1, characterized in that, In step 204, according to the preset momentum decay factor, the first The momentum term after the second iteration update is the same as the first iteration. The gradient of the nth iteration is updated through momentum accumulation to obtain the th... The momentum term updated in the next iteration is expressed as follows: ； In the formula, Indicates the first The momentum term updated in the next iteration; Indicates the momentum decay factor; Indicates the first The momentum term updated in the next iteration; Indicates the first The gradient of the next iteration; Indicates the first Loss per iteration; This represents hyperspectral data containing perturbations; Represents a cube containing the raw hyperspectral data; Indicates the first The test perturbation after the next iteration optimization; express Norm.

5. The robustness evaluation method for depth reconstruction models in hyperspectral computational imaging according to claim 1, characterized in that, In step 205, according to the first The momentum term updated in the next iteration, the first The test perturbation after the second iteration optimization is adjusted along the momentum term direction and clipped to the constraint range, with the preset iteration step size and maximum perturbation amplitude as constraints, and the result is the output of the second iteration. The test perturbation after the next iteration of optimization is expressed as follows: ； In the formula, Indicates the first The test perturbation after the next iteration optimization; Represents the clipping function; Indicates the first The test perturbation after the next iteration optimization; Indicates the preset iteration step size; Represents a symbolic function; Represents a cube containing the raw hyperspectral data; Indicates the maximum amplitude of the disturbance; Indicates the first The momentum term after the next iteration update.

6. The robustness evaluation method for depth reconstruction models of hyperspectral computational imaging according to any one of claims 1 to 5, characterized in that, In step 201, if the target hyperspectral computational imaging system is a coded aperture snapshot spectral imaging system, the mask enhancement distribution is a uniform mixture of a binary mask distribution and a grayscale mask distribution; Specifically, a basic binary mask matrix is ​​generated by sampling the transmittance parameter; then, the basic binary mask matrix is ​​upsampled to the target space size to obtain the binary mask distribution. A basic grayscale mask matrix is ​​generated by sampling the mean and standard deviation; the basic grayscale mask matrix is ​​then cropped, and the cropped basic grayscale mask matrix is ​​upsampled to the target space size to obtain the grayscale mask distribution.

7. The robustness evaluation method for depth reconstruction models of hyperspectral computational imaging according to any one of claims 1 to 5, characterized in that, In step 201, if the target hyperspectral computational imaging system is an imaging system with an optical filter, the mask enhancement distribution is a three-dimensional grayscale mask distribution. Specifically, a basic three-dimensional mask tensor is generated by sampling the mean and standard deviation; the basic three-dimensional mask tensor is clipped, and the clipped basic three-dimensional mask tensor is upsampled to the target space and spectral size to obtain a three-dimensional grayscale mask distribution.

8. The robustness evaluation method for depth reconstruction models in hyperspectral computational imaging according to claim 1, characterized in that, Step 3 also includes: For the final test sample, the corresponding evaluation compressed measurement is generated by encoding the physical forward model of the target hyperspectral computational imaging system. The evaluation compression measurement is input into the depth reconstruction model to be evaluated to generate the reconstruction result.

9. A robustness evaluation device for a depth reconstruction model of hyperspectral computational imaging, characterized in that, The apparatus for evaluating the robustness of a depth reconstruction model for hyperspectral computational imaging according to any one of claims 1 to 8 comprises: The data acquisition module is used to acquire the raw hyperspectral data cube and target physical mask to be evaluated; The iterative optimization module is used to generate an optimized test perturbation based on the initial test perturbation and initial momentum term, using a momentum iterative optimization framework and incorporating a mask enhancement strategy, through multiple iterations. The test sample generation module is used to apply the optimized test perturbation to the original hyperspectral data cube to generate the final test sample. The evaluation module is used to input the final test sample into the depth reconstruction model to be evaluated, generate reconstruction results, and quantify the robustness of the evaluation model by comparing the differences between the reconstruction results and the evaluation indicators of the original hyperspectral data cube.

Images