Design method, system, and storage medium for hyperspectral computed imaging encoders

Abstract

The application is suitable for the technical field of optical filters, and discloses a design method of a hyperspectral computational imaging encoder. The application adopts a cooperative optimization architecture of a main population and an auxiliary population, takes the thickness of each layer of the thin film of the filter array as a design variable, generates candidates through layered differential search, total thickness adjustment guided by transmittance, and channel replacement alignment reorganization operators; combines ridge regression proxy pre-screening and optical simulation evaluation by the transfer matrix method, guarantees feasible convergence with strict constraints for the main population, expands the search space with dynamic relaxation constraints for the auxiliary population, and obtains a Pareto optimal solution set through cooperative iteration of the two, and finally selects the optimal design based on the reconstructed peak signal-to-noise ratio. Compared with the prior art, the application effectively improves the global optimization ability and convergence efficiency, considers constraint satisfaction, multi-objective optimization and manufacturing robustness, and can obtain a hyperspectral computational imaging encoder with better reconstruction performance.

Description

Technical Field

[0001] This invention relates to the field of filter technology, and in particular to a design method, system and storage medium for a hyperspectral computational imaging encoder. Background Technology

[0002] Hyperspectral imaging technology can simultaneously acquire two-dimensional spatial information and continuous spectral information of a target scene, and has significant application value in many fields such as resource remote sensing, biomedical detection, industrial non-destructive sorting, and agricultural product quality analysis. The computational hyperspectral imaging architecture based on a multi-channel filter array encodes and modulates the incident spectrum through a filter array with differentiated spectral transmittance characteristics. This is then combined with a back-end spectral reconstruction algorithm to obtain full-band hyperspectral data, enabling snapshot-like imaging. It also boasts advantages such as compact system structure, small size, and light weight, making it the mainstream technology for miniaturized and integrated hyperspectral imaging systems. Among these technologies, the filter array, as the core component of spectral encoding, directly determines the spectral reconstruction accuracy and imaging quality of the entire imaging system through the encoding capability of its transmittance curve combination. Therefore, the optimized design of the filter array is a crucial aspect of hyperspectral computational imaging systems.

[0003] Currently, the design methods for hyperspectral filter arrays mainly fall into two categories. One is a filter selection framework based on genetic algorithms, which constructs a pre-defined candidate library of spectral transmittance curves and uses reconstruction accuracy as the fitness index to select the optimal filter combination from the candidate library using an evolutionary algorithm. The other is an end-to-end collaborative optimization framework based on deep learning, which jointly trains the spectral encoding module and the spectral reconstruction module, and uses reconstruction error as the loss function to iteratively optimize the structural parameters of the filters.

[0004] However, both of these design paradigms have significant technical limitations.

[0005] First, the design results are strongly coupled with specific reconstruction algorithms, resulting in insufficient universality. Most existing methods use the reconstruction metrics of specific reconstruction algorithms (such as relative root mean square error, peak signal-to-noise ratio, and structural similarity) as optimization targets, and the optimal solution obtained is only suitable for the corresponding reconstruction algorithm. When the backend is changed to a reconstruction algorithm with different principles, the coding performance of the filter array will be significantly degraded, and it cannot be universally applied across algorithms.

[0006] Secondly, the optimization process is highly dependent on the training dataset. Optimization driven by reconstruction results requires a large amount of computing power for reconstruction calculations, and the design results are significantly affected by the size and distribution of the training data. When the amount of training data is insufficient or the distribution is unbalanced, the performance deviation of the reconstruction algorithm itself will be transmitted to the filter design stage, resulting in the final selected scheme not being a reasonable coding structure.

[0007] Furthermore, design patterns based on candidate library selection are subject to design space boundary limitations. Most existing genetic algorithm schemes can only perform combination screening within a pre-given, limited curve library, and cannot explore non-classical membrane structures outside the candidate library. The optimality of the design results is limited by the coverage of the candidate library, making it difficult to obtain a higher-performance encoding scheme.

[0008] More critically, existing filter design methods are generally based on the idealized assumption that "design value is manufacturing value," always evaluating and optimizing performance under nominal structural parameters, without taking into account the objective process errors that exist in micro- and nano-fabrication. This leads to a significant performance gap between theoretical design and actual devices. For multilayer dielectric thin-film filters, the actual deposition thickness of each layer inevitably deviates from the design value by nanometers due to factors such as the control precision of the coating equipment, fluctuations in deposition uniformity, substrate temperature drift, and long-term deposition accumulation errors. Furthermore, the spectral response of multilayer films is determined by the coherent superposition of reflected waves from each layer interface, making it highly sensitive to phase changes. Small deviations in the thickness of a single layer will accumulate and amplify layer by layer, ultimately causing problems such as transmission peak shift, bandwidth distortion, and modulation distortion. At the same time, the nominal optimal solution in a high-dimensional non-convex design space is often extremely sensitive to parameter perturbations. Even small process deviations can lead to a sharp decline in performance, causing designs with excellent on-paper performance to experience significant performance degradation after fabrication. While some existing studies have identified the sensitivity of ultrathin films to process errors, they have not systematically incorporated processing errors into the optimization design process, thus failing to address the root cause of insufficient manufacturing robustness in the design schemes.

[0009] In summary, existing design methods for hyperspectral computational imaging encoders struggle to simultaneously achieve high coding performance, cross-algorithm universality, and manufacturing robustness, thus hindering the performance improvement and industrialization of hyperspectral computational imaging systems.

[0010] Therefore, there is an urgent need for a new design method, system, and storage medium for hyperspectral computational imaging encoders to solve the above-mentioned technical problems. Summary of the Invention

[0011] This invention provides a design method, system, and storage medium for a hyperspectral computational imaging encoder, aiming to solve the technical problems of insufficient process robustness, multi-objective performance imbalance, poor engineering manufacturability, and premature convergence in optimization of existing hyperspectral computational imaging encoder design methods. It effectively improves the process robustness, overall coding performance, engineering manufacturability, and optimization efficiency of the hyperspectral computational imaging encoder design scheme.

[0012] In a first aspect, the present invention provides a design method for a hyperspectral computational imaging encoder, the design method comprising the following steps:

[0013] S1. Using the thickness of each thin film layer of the filter array as the design variable, construct an initial primary candidate population and an initial auxiliary candidate population based on a hybrid initialization strategy. Modify the initial primary candidate population and the initial auxiliary candidate population respectively through manufacturing feasibility constraints to obtain a legal primary population and a legal auxiliary population. The initial primary candidate population and the initial secondary candidate population each include multiple genetic individuals, with each genetic individual corresponding to a group. M×N L The filter thickness matrix, M The number of filters in the filter array. N L The number of thin film layers in a single filter in the filter array, wherein each filter comprises sequentially arranged... N L One gene unit, N L Each of the aforementioned gene units corresponds to a filter layer from the first layer to the second layer. N L The thickness of the thin film layer; S2. Calculate the budget progress based on the current consumed evaluation budget and the preset evaluation budget limit, and perform an overall stagnation detection on the legitimate main population to obtain the stagnation detection result; select the corresponding strategy based on the budget progress and the stagnation detection result, and calculate the mating fitness values ​​corresponding to the legitimate main population and the legitimate auxiliary population. S3. Perform hierarchical differential search processing on the legitimate primary population and the legitimate auxiliary population according to the mating fitness value to obtain the offspring population; S4. The offspring population is screened based on the ridge regression surrogate model, and optical simulation is performed on each genetic individual in the screened offspring population using the transfer matrix method to obtain the spectral modulation matrix of the corresponding filter array; an optimization target is established, and the nominal performance value corresponding to the optimization target is calculated based on the spectral modulation matrix. Based on performance constraints, each genetic individual in the offspring population is verified and labeled, and the consumed evaluation budget is updated to obtain a constraint feasibility label. S5. Based on the nominal performance value and the constraint feasibility marker, perform environmental selection on the offspring population to obtain the next generation main population and the next generation auxiliary population; wherein, when selecting the next generation main population, the selection is based on a preset boundary from the legal main population and the offspring population; when selecting the next generation auxiliary population, the selection is based on a dynamic boundary from the legal auxiliary population, the offspring population and the next generation main population; S6. Determine whether the updated consumed evaluation budget has reached or exceeded the preset evaluation budget limit: If so, extract all non-dominated solutions from the current next-generation master population to obtain the Pareto optimal filter array design solution set, calculate the peak signal-to-noise ratio corresponding to each candidate design in the Pareto optimal filter array design solution set, and output the candidate design with the largest peak signal-to-noise ratio as the final design scheme of the hyperspectral computational imaging encoder. If not, then the next generation main population is taken as the new legitimate main population, and the next generation auxiliary population is taken as the new legitimate auxiliary population, and the process returns to step S2 to continue iterative optimization.

[0014] Preferably, in step S1, the hybrid initialization strategy is as follows: the film thickness of the filter for the genetic individual is generated through constrained random initialization, quarter-wavelength optical thickness initialization, and Fabry-Perot cavity initialization. The constraint is randomly initialized as follows: the total thickness value of a single filter is randomly generated within a preset total thickness range, and the total thickness value is proportionally distributed to each layer of the filter film using random weights. The quarter-wavelength optical thickness is initialized by randomly selecting the center wavelength, calculating the thickness of each thin film layer of the filter based on the quarter-wavelength interference condition, and applying multiplicative Gaussian noise perturbation. The Fabry-Perot cavity is initialized by amplifying the film thickness value corresponding to the intermediate layer of the filter based on the quarter-wavelength basic film system structure and applying multiplicative noise perturbation.

[0015] Preferably, the manufacturing feasibility constraints include material system constraints, structural size constraints, single-layer thickness constraints, and total thickness constraints: The material system constraint is that the filter adopts a preset high and low refractive index material system and a fixed alternating stacking order; The structural size constraint is that the number of thin film layers of the filter and the total number of channels in the filter array are both preset fixed values; The single-layer thickness constraint means that the film thickness of each layer of the filter must be within the preset upper and lower limits of the single-layer thickness. The total thickness constraint means that the total thickness of the filter must be within a preset total thickness target range.

[0016] Preferably, step S2 further includes the following sub-steps: S21. Establish an internal generation clock, calculate the dynamic boundary based on the internal generation clock, and update the boundary of the legitimate auxiliary population based on the dynamic boundary; wherein, the internal generation clock is dynamically adjusted according to the number of population iterations; S22. Perform overall stagnation detection on the legitimate main population to obtain stagnation detection results; calculate the budget progress based on the current consumed evaluation budget and the preset evaluation budget upper limit; S23. When the stagnation detection result is stagnation or the budget progress exceeds the preset budget progress threshold, the mating fitness values ​​corresponding to the legitimate primary population and the legitimate auxiliary population are calculated based on the robust evaluation strategy. When the stagnation detection result is no stagnation or the budget progress does not exceed the preset budget progress threshold, the mating fitness values ​​corresponding to the legitimate primary population and the legitimate secondary population are calculated using the Pareto dominance strategy.

[0017] Preferably, the robustness evaluation strategy is as follows: The legitimate primary population and the legitimate secondary population are deduplicated and merged to obtain a mixed population; The convergence of each genetic individual in the mixed population is quantified using an achievement scalarization function, resulting in multiple convergence values. The achievement scalarization function satisfies the following condition: ; in, Indicates genetic individual x In the The target value on each target, This indicates that all current genetic individuals are in the [number]th [year]. The optimal value for each objective. Indicates genetic individual x In the r A preference weight vector on each objective The achievement scalarization function is represented by the convergence value obtained by normalizing the result of the achievement scalarization function. ; A manufacturing error perturbation is applied to the genetic individuals of the mixed population to obtain a perturbation target value. The sum of the deviations of the genetic individuals of the mixed population from the perturbation target value is calculated as a robustness value. The robustness value satisfies the following condition:

[0018] ; in, This represents the robustness value. Indicates genetic individual x In the r Deviation from the target Indicates genetic individual x After the disturbance The perturbation target value on each target; A comprehensive performance index is calculated based on the robustness value and the convergence value, and this comprehensive performance index is used as the mating fitness value. The comprehensive performance index satisfies the following conditions: ; ; in, Indicates adaptive weights, This represents the median of the achievement scalarization function results for all genetically derived individuals in the mixed population. and This represents the historical maximum and minimum values ​​of the achievement scalarization function during the iteration process. Indicates genetic individual x Corresponding comprehensive performance indicators; The Pareto domination strategy is as follows: domination relationships are determined for all genetic individuals in the mixed population pairwise, and a constraint priority rule is introduced in the domination determination process; wherein, the constraint priority rule is: when comparing two genetic individuals, the genetic individual with a smaller total constraint violation has a higher priority. The number of other individuals dominated by each genetically derived individual is counted to obtain the intensity value; The density score is obtained by calculating the neighbor distances of each genetic individual in the target space corresponding to the optimization objective to the neighboring individuals. The comprehensive score of each genetic individual is calculated based on the intensity value and the density score, and the comprehensive score is used as the mating fitness value.

[0019] Preferably, step S4 includes the following sub-steps: S41. For each genetic individual in the offspring population, forward simulation is performed using the transfer matrix method to obtain the spectral transmittance curves of each filter channel in the full band. Multiple spectral transmittance curves are spliced ​​together to obtain the spectral modulation matrix. S42. Calculate the nominal performance value corresponding to the optimization objective based on the spectral modulation matrix; wherein, the optimization objective includes an information volume measurement objective, a spectral response structure independence objective, and an inverse problem benignity objective: The information volume measurement objective is: based on the Bayesian D-optimality criterion, to measure the total amount of information captured from the scene by the filter array represented by the spectral modulation matrix, and to drive the optimization of the light flux, channel independence and band targeting in the corresponding dimension of the spectral modulation matrix; The objective of spectral response structure independence is to measure the non-redundancy of the encoded information of each filter channel of the spectral modulation matrix by decentering the spectral response curves of each filter channel of the spectral modulation matrix and introducing a wavelength offset. The benignity objective of the inverse problem is: the degree of amplification of the noise in the inversion and reconstruction after the observation signal is encoded by the spectral modulation matrix, which is quantified by the conditions of the spectral modulation matrix; S43. Verify whether each genetic individual in the offspring population satisfies the performance constraint, mark genetic individuals that satisfy the performance constraint as feasible individuals, and mark genetic individuals that do not satisfy the performance constraint as infeasible individuals, thus obtaining constraint feasibility labels; wherein, the ranking priority of any feasible individual is higher than that of all infeasible individuals, and the performance constraint is to make the offspring population satisfy the following conditions: The average transmittance of all filter channels in the offspring population is not lower than a preset transmittance threshold. The standard deviation of the average transmittance among the filter channels in the offspring population is less than a preset standard deviation threshold. The spectral resolution of all filter channels in the offspring population is not lower than the preset spectral resolution.

[0020] Preferably, step S3 includes the following sub-steps: S31. Based on the binary tournament mechanism, individuals are extracted from the legitimate main population to obtain a first extracted parent main population. The first extracted parent main population is sorted in ascending order by the mating fitness and split into a superior subset and a inferior subset. Diversity difference is performed on the superior subset, and convergence difference is performed on the inferior subset to obtain the difference result. Boundary pruning and polynomial mutation processing are performed on the difference result to obtain the main population difference subset. S32. Perform the same processing logic as in step S31 on the legal auxiliary population to obtain the auxiliary population difference numerator; S33. Random individuals are randomly extracted from the legal main population to obtain a second extracted parent main population; the average channel film thickness of each channel in the second extracted parent main population is adjusted according to the normalized average transmittance constraint residual corresponding to the second extracted parent main population, and a perturbation is applied to the film thickness of a random number of channels in the adjusted second extracted parent main population to obtain a thickness adjustment subset. S34. Random individuals are randomly extracted from the legitimate main population to obtain a baseline parent population; individuals are uniformly and randomly extracted from the legitimate main population and the legitimate auxiliary population to obtain a donor population; a cost matrix is ​​constructed between the baseline parent population and the donor population using the Euclidean distance of the channel thickness vector; and the minimum cost pair that has not yet been occupied in the cost matrix is ​​repeatedly selected based on a greedy algorithm; the channels of the baseline parent population are mapped and aligned one by one with the channels of the donor population; a random number of channels in the baseline parent population are selected to perform whole channel replacement or layer thickness convex combination interpolation and recombination and perturbation is applied to obtain an aligned recombination subset; S35. Merge the main population difference subset, the auxiliary population difference subset, the thickness adjustment subset, and the alignment recombination subset to obtain the candidate offspring population; S36. The candidate offspring population is modified by the manufacturing feasibility constraints and candidate deduplication is performed to obtain the offspring population.

[0021] Secondly, the present invention also provides a design system for a hyperspectral computational imaging encoder, the design system comprising: The initialization module is used to construct an initial primary candidate population and an initial auxiliary candidate population based on a hybrid initialization strategy, using the thickness of each thin film layer of the filter array as a design variable. The initial primary candidate population and the initial auxiliary candidate population are then modified by manufacturing feasibility constraints to obtain a legal primary population and a legal auxiliary population. The initial primary candidate population and the initial secondary candidate population each include multiple genetic individuals, with each genetic individual corresponding to a group. M×N L The filter thickness matrix, M The number of filters in the filter array. N L The number of thin film layers in a single filter in the filter array, wherein each filter comprises sequentially arranged... N L One gene unit, N L Each of the aforementioned gene units corresponds to a filter layer from the first layer to the second layer. N L The thickness of the thin film layer; The mating fitness calculation module is used to calculate the budget progress based on the current consumed evaluation budget and the preset evaluation budget limit, and to perform overall stagnation detection on the legitimate main population to obtain the stagnation detection result; select the corresponding strategy based on the budget progress and the stagnation detection result, and calculate the mating fitness values ​​corresponding to the legitimate main population and the legitimate auxiliary population; The hierarchical differential module is used to perform hierarchical differential search processing on the legitimate primary population and the legitimate auxiliary population respectively according to the mating fitness value to obtain the offspring population; The optimization target module is used to screen the offspring population based on the ridge regression surrogate model, and to perform optical simulation on each genetic individual in the screened offspring population using the transfer matrix method to obtain the spectral modulation matrix of the corresponding filter array; to establish the optimization target, and to calculate the nominal performance value corresponding to the optimization target based on the spectral modulation matrix; Based on performance constraints, each genetic individual in the offspring population is verified and labeled, and the consumed evaluation budget is updated to obtain a constraint feasibility label. A screening module is used to perform environmental selection on the offspring population based on the nominal performance value and the constraint feasibility marker, thereby selecting the next generation main population and the next generation auxiliary population; wherein, when screening the next generation main population, the selection is based on a preset boundary from the legal main population and the offspring population; when screening the next generation auxiliary population, the selection is based on a dynamic boundary from the legal auxiliary population, the offspring population, and the next generation main population; The judgment module is used to determine whether the updated consumed evaluation budget has reached or exceeded the preset evaluation budget limit: If so, extract all non-dominated solutions from the current next-generation master population to obtain the Pareto optimal filter array design solution set, calculate the peak signal-to-noise ratio corresponding to each candidate design in the Pareto optimal filter array design solution set, and output the candidate design with the largest peak signal-to-noise ratio as the final design scheme of the hyperspectral computational imaging encoder. If not, then the next generation primary population is taken as the new legitimate primary population, and the next generation secondary population is taken as the new legitimate secondary population, and the process returns to the mating fitness calculation module to continue iterative optimization.

[0022] Thirdly, the present invention also provides a computer device, comprising: a memory, a processor, and a design program for a hyperspectral computational imaging encoder stored in the memory and executable on the processor, wherein when the processor executes the design program for the hyperspectral computational imaging encoder, it implements the steps in the design method of the hyperspectral computational imaging encoder as described in any of the above embodiments.

[0023] Fourthly, the present invention also provides a computer-readable storage medium storing a design program for a hyperspectral computational imaging encoder, wherein when the design program for the hyperspectral computational imaging encoder is executed by a processor, it implements the steps in the design method of the hyperspectral computational imaging encoder as described in any of the above embodiments.

[0024] Compared with existing technologies, this invention constructs a multi-dimensional evaluation system of optimization objectives based on the inherent physical characteristics of the spectral modulation matrix. This system consists of an information volume measurement objective, a spectral response structure independence objective, and an inverse problem benignity objective. The nominal performance value is used as the basis for optimization. Performance evaluation and optimization can be completed without relying on specific spectral reconstruction algorithms and training datasets. The resulting filter array scheme is not affected by changes in the backend reconstruction algorithm and has cross-algorithm versatility. At the same time, it avoids the computational power consumption caused by a large number of reconstruction iteration calculations, and significantly improves design efficiency.

[0025] This invention directly uses the thickness of each thin film layer in the filter thickness matrix as a design variable, and generates a new film system structure through an evolutionary algorithm, rather than combining and screening from a preset candidate library. It can explore non-classical structural forms outside of traditional classic film systems, break the constraints of the candidate library on the design space, and obtain filter array schemes with stronger encoding capabilities and better spectral modulation characteristics.

[0026] This invention uses three mutually balancing performance indicators as parallel optimization objectives and obtains a Pareto-optimal filter array design solution set composed of non-dominated solutions through multi-objective evolutionary search. By postponing the engineering trade-off step to the end of the optimization process, the final solution can be selected from the leading solution set according to the performance emphasis of different application scenarios, thus avoiding performance bias caused by subjective weighted summation.

[0027] This invention employs a hybrid initialization strategy combining constrained random initialization, quarter-wavelength optical thickness initialization, and Fabry-Perot cavity initialization to balance the structural rationality and morphological diversity of the initial population. It utilizes crossover operations with the complete thickness vector corresponding to a single filter as the crossover unit, combined with sparse adaptive Gaussian mutation and random replacement mutation, to improve evolutionary efficiency while preserving the effective interlayer coordination structure. Furthermore, it achieves adaptive environment selection by integrating a comprehensive performance index based on diversity evaluation of the target space sub-region division and convergence and robustness values ​​calculated using a scalarization function. This dynamically adjusts the exploration direction according to the population state, balancing convergence speed and population diversity.

[0028] At the constraint and control level, the initial candidate population and offspring candidate population are modified in real time by manufacturing feasibility constraints, and legal initial population and legal offspring population are output throughout the process. Combined with the dominance mechanism of performance soft constraints and constraint feasibility markers, it ensures that all candidate solutions are within the process feasible domain and meet the basic performance requirements such as transmittance, uniformity, and spectral resolution, thus avoiding the generation of invalid solutions with no engineering application value.

[0029] This invention employs strict constraint screening on the main population, focusing on multi-objective convergence within the feasible region to ensure the constraint compliance and Pareto front quality of the final solution. The auxiliary population uses dynamic relaxation constraints, retaining individuals that are temporarily infeasible but possess excellent target performance and exploration value, effectively expanding the search space, maintaining population diversity, and helping the algorithm traverse infeasible regions and escape local optima. Both populations jointly generate candidate solutions, and the offspring of truly evaluated solutions participate in the dual-population update simultaneously. Simultaneously, the updated high-quality solutions from the main population are directionally injected into the auxiliary population, providing feasibility guidance for the exploration process and preventing the search from deviating from the feasible direction, thereby achieving multi-directional collaboration and improved efficiency. The overall architecture achieves a synergistic balance between global exploration, constraint satisfaction, and multi-objective convergence. While ensuring the design meets manufacturing and performance constraints, it effectively improves the algorithm's global optimization capability and convergence efficiency, resulting in a hyperspectral computational imaging encoder design with superior reconstruction performance. Attached Figure Description

[0030] The present invention will now be described in detail with reference to the accompanying drawings. The above and other aspects of the present invention will become clearer and more readily understood through the detailed description following the accompanying drawings. In the drawings: Figure 1 This is a flowchart of the design method for a hyperspectral computational imaging encoder provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the spectral transmittance curve of the filter array obtained by forward simulation using the transfer matrix method according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the design system of the hyperspectral computational imaging encoder provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] Example 1 Please refer to Figure 1 This invention provides a design method for a hyperspectral computational imaging encoder, the design method comprising the following steps: S1. Using the thickness of each thin film layer of the filter array as the design variable, construct an initial primary candidate population and an initial auxiliary candidate population based on a hybrid initialization strategy. Modify the initial primary candidate population and the initial auxiliary candidate population respectively through manufacturing feasibility constraints to obtain a legal primary population and a legal auxiliary population. The initial primary candidate population and the initial secondary candidate population each include multiple genetic individuals, with each genetic individual corresponding to a group. M×N L The filter thickness matrix, M The number of filters in the filter array. N L The number of thin film layers in a single filter in the filter array, wherein each filter comprises sequentially arranged... N L One gene unit, N L Each of the aforementioned gene units corresponds to a filter layer from the first layer to the second layer. N L The thickness of the thin film layer.

[0033] In this embodiment of the invention, in step S1, the hybrid initialization strategy is as follows: the film thickness of the filter for the genetic individual is generated through constrained random initialization, quarter-wavelength optical thickness initialization, and Fabry-Perot cavity initialization. The constraint is randomly initialized as follows: the total thickness value of a single filter is randomly generated within a preset total thickness range, and the total thickness value is proportionally distributed to each layer of the filter film using random weights. The quarter-wavelength optical thickness is initialized by randomly selecting the center wavelength, calculating the thickness of each thin film layer of the filter based on the quarter-wavelength interference condition, and applying multiplicative Gaussian noise perturbation. The Fabry-Perot cavity is initialized by amplifying the film thickness value corresponding to the intermediate layer of the filter based on the quarter-wavelength basic film system structure and applying multiplicative noise perturbation.

[0034] The manufacturing feasibility constraints include material system constraints, structural size constraints, single-layer thickness constraints, and total thickness constraints: The material system is constrained by the fact that the filter adopts a preset high and low refractive index material system and a fixed alternating stacking order; specifically, odd-numbered layers are made of low refractive index materials, and even-numbered layers are made of high refractive index materials, which alternate to form an interference film stack, and the substrate is optical glass with high transmittance.

[0035] The structural size constraint is that the number of thin film layers of the filter and the total number of channels in the filter array are both preset fixed values; The single-layer thickness constraint means that the film thickness of each layer of the filter must be within the preset upper and lower limits of the single-layer thickness. The total thickness constraint means that the total thickness of the filter must be within a preset total thickness target range.

[0036] S2. Calculate the budget progress based on the current consumed evaluation budget and the preset evaluation budget limit, and perform an overall stagnation detection on the legitimate main population to obtain the stagnation detection result; select the corresponding strategy based on the budget progress and the stagnation detection result, and calculate the mating fitness values ​​corresponding to the legitimate main population and the legitimate auxiliary population.

[0037] In this embodiment of the invention, step S2 further includes the following sub-steps: S21. Establish an internal generation clock, calculate the dynamic boundary based on the internal generation clock, and update the boundary of the legitimate auxiliary population based on the dynamic boundary; wherein, the internal generation clock is dynamically adjusted according to the number of population iterations.

[0038] Specifically, definition For the current consumed evaluation budget, To preset the upper limit of the evaluation budget, N This represents the configuration capacity corresponding to the main population and the auxiliary population, and is also the number of reference vectors and the maximum true evaluation quota of the regular generation. Indicates genetic individual The target value vector of multiple optimization objectives; Indicates genetic individual Multiple constraint residual value vectors, Representing the The genetic individual satisfies the first... One constraint Indicates genetic individual The total number of constraint violations satisfies: .

[0039] Internal Generation Clock Starting from 0, each generation increments by the condition that satisfies... = Up to the cap of 1, it only serves the tightening rhythm of the dual boundaries. Because the most genuine reviews are from the regular first generation. A new design, The growth rate is approximately It is twice the size of the two, but they are not completely synchronized and need to be distinguished during use.

[0040] The dynamic boundary includes a first boundary and a second boundary, which are two thresholds designed as a dynamic trade-off between "tolerating violations and encouraging early exploration" and "approaching the truly feasible region," and tighten with evolution. First, a violation scale benchmark is defined on the joint set of the two initial populations. (If all initial individuals are feasible) To avoid subsequent occurrences These meaningless operations, forced The boundary tightening index is defined as:

[0041] Its motivation is to make the first boundary Exactly at the "halfway point" of the internal clock ( decay to This is a... An irrelevant absolute anchor point, rather than The contraction factor relative to the initial violation scale is the only factor. In other words, regardless of the initial scale of violations, the tolerable level of violations is suppressed in the medium term. This fixed absolute magnitude allows sufficient room for tightening in later approximations of the feasible region. This formula is only valid when... Only then can it be guaranteed Follow Monotonous decline. Determine the width of the "reference set" in environment selection: the violation amount shall not exceed Individuals selected are included in the reference set and used as a benchmark for comparison with the remaining individuals.

[0042] Second boundary When determining whether constraints need to be considered when comparing indicators, the following should be considered: ; The proportion of approximately feasible individuals in the main population ( It is a tolerance that absorbs numerical residuals, which is different from the exact zero violation used in reference set determination. The larger the value, the closer the main population is to complete feasibility. In this case, the comparison should be more strictly performed according to the constraints. The population size subsequently narrows or even approaches zero; the main population is entirely feasible. )hour Any challenger who violates the rules must simultaneously undergo a constraint comparison. The two populations use different strictness boundaries: the dominant population always uses strict boundaries during environmental selection. Only truly feasible solutions can be included in the reference set; if no feasible solution can be found, the process degenerates into a system based on... The first half of the ascending sequence is used as the alternative reference set; the auxiliary population uses the above dynamics. This allows individuals with low violation rates to be temporarily treated as "promising" in the early stages, preserving more feasible space for extraterritorial exploration in the search.

[0043] S22. Perform an overall stagnation detection on the legitimate main population to obtain the stagnation detection result; calculate the budget progress based on the current consumed evaluation budget and the preset evaluation budget upper limit.

[0044] Among them, the evaluation budget progress meets the following requirements: It reflects the actual consumption ratio of evaluation resources, thereby driving the intensity scheduling of robust switching criteria and search operators.

[0045] Maintain a one-way state while updating the double boundary. This is used to determine whether the legitimate primary population objective has stagnated as a whole. Each iteration sets the empirical ideal point of the legitimate primary population objective... The worst experience (Maximum value per target, not the nadir point in the strict Pareto sense) and mean Added to the trajectory record. These three types of statistics each have their own emphasis: the ideal point describes the best edge (convergence front) that each target can currently reach, the worst point describes the expansion range of the population on each target, and the mean point describes the overall centroid of the target distribution; according to this criterion, all three must be frozen at the same time to be considered as stagnation, which is more conservative and comprehensive than only looking at the convergence of the ideal point.

[0046] Where, the lower limit of the denominator To avoid scaling distortion when historical values ​​are close to zero. Here. It is a placeholder symbol, generally referring to the first position in the trajectory record. The value of a scalar statistical component in the current generation, with its subscript. For serial numbers, This indicates the value of the same component at the starting point of the window; for In terms of objectives, the three types of statistics each Dimensions, Total Each scalar component, relative change Each of these nine components needs to be calculated individually, so its essence is a component-by-component (vector) quantity. Ultimately, it is reduced to a single scalar by taking the maximum value and then compared with a threshold (using the maximum value rather than the average). If the maximum value of all relative changes does not exceed... (default Approximately corresponding If the threshold for change is reached, then stagnation is determined. Change from 1 to 2 in one direction.

[0047] when And the cumulative records reach the window length (Default 20, meaning spanning 19 generation intervals) When comparing the current record with... For previous records, calculate the relative change for each objective of the three types of statistics: ; The robustness evaluation is delayed by using a stagnant state. The robustness evaluation is determined by problem parameters. control: (Default) is purely nominal evaluation; And algorithm parameters (By default) a "nominal first, robust later" scheduling strategy is adopted, and at startup, the problem is first allowed to run in Pareto dominance mode. Initialize and search until the switching conditions are met: ; Either of these events triggers an immediate switch; even if a standstill hasn't actually been detected, the half-term budget safety net will force a switch, and They will all be locked at level 2, and the stall detection will no longer proceed. Restore during switching. A formal robustness evaluation should be re-performed for all individuals in both populations, and mating fitness should be recalculated. If... Robust evaluation is used throughout the entire process from initialization. Robust evaluation applies manufacturing error perturbations to each nominal design and uses a comprehensive performance index defined below, based on the Achievement Scalar Function (ASF). As a robustness assessment result, the constraints are always taken as the nominal design values. That is, the constraints describe whether the design specifications themselves are met, and robustness is only applied to optical performance targets that are directly affected by process errors.

[0048] S23. When the stagnation detection result is stagnation or the budget progress exceeds the preset budget progress threshold, the mating fitness values ​​corresponding to the legitimate primary population and the legitimate auxiliary population are calculated based on the robust evaluation strategy. When the stagnation detection result is no stagnation or the budget progress does not exceed the preset budget progress threshold, the mating fitness values ​​corresponding to the legitimate primary population and the legitimate secondary population are calculated using the Pareto dominance strategy.

[0049] In this embodiment of the invention, the robustness evaluation strategy is as follows: The legitimate primary population and the legitimate secondary population are deduplicated and merged to obtain a mixed population; The convergence of each genetic individual in the mixed population is quantified using an achievement scalarization function, resulting in multiple convergence values. The achievement scalarization function satisfies the following condition: ; in, Indicates genetic individual x In the The target value on each target, This indicates that all current genetic individuals are in the [number]th [year]. The optimal value for each objective. Indicates genetic individual x In the r A preference weight vector on each objective The achievement scalarization function is represented by the convergence value obtained by normalizing the result of the achievement scalarization function. ; A manufacturing error perturbation is applied to the genetic individuals of the mixed population to obtain a perturbation target value. The sum of the deviations of the genetic individuals of the mixed population from the perturbation target value is calculated as a robustness value. The robustness value satisfies the following condition:

[0050]

[0051] in, This represents the robustness value. Indicates genetic individual x In the r Deviation from the target Indicates genetic individual x After the disturbance The perturbation target value on each target; A comprehensive performance index is calculated based on the robustness value and the convergence value, and this comprehensive performance index is used as the mating fitness value. The comprehensive performance index satisfies the following conditions: ; ; in, Indicates adaptive weights, This represents the median of the achievement scalarization function results for all genetically derived individuals in the mixed population. and This represents the historical maximum and minimum values ​​of the achievement scalarization function during the iteration process. Indicates genetic individual x Corresponding comprehensive performance indicators; Indicates genetic individual The corresponding comprehensive performance index (by convention, the smaller the better). This adaptive weight uses the relative position of the current ASF median in the historical range as the convergence weight: when the overall ASF of the population is high (not yet converged) Larger, with a greater emphasis on convergence As iterations approach the frontier and the overall ASF decreases... Reduce, and relatively emphasize robustness .

[0052] The Pareto domination strategy is as follows: domination relationships are determined for all genetic individuals in the mixed population pairwise, and a constraint priority rule is introduced in the domination determination process; wherein, the constraint priority rule is: when comparing two genetic individuals, the genetic individual with a smaller total constraint violation has a higher priority. The number of other individuals dominated by each genetically derived individual is counted to obtain the intensity value; The density score is obtained by calculating the neighbor distances of each genetic individual in the target space corresponding to the optimization objective to the neighboring individuals. The comprehensive score of each genetic individual is calculated based on the intensity value and the density score, and the comprehensive score is used as the mating fitness value.

[0053] Specifically, the overall score is based on the SPEA2 algorithm's approach of dominance counting and density estimation: first, define the dominance relationship with constrained priorities— hour Dominate Only when Only then does it revert to the standard Pareto dominance comparison target vector. Let... For genetic individuals Number of dominant individuals, initial fitness Accumulate the dominance counts of all dominators (a larger value indicates a more crowded and worse region); the density term is taken from the original target space. Euclidean distance between nearest neighbors ,make (Constant 2 avoids zero-distance singularity and puts) Limited to (To prevent the density term from overshadowing the main term when the dominance counts are equal). Final score Smaller is better; It is a sublinear trade-off commonly used between locality and noise resistance. The distance is not normalized for each target, so targets with larger numerical ranges will have a greater weight in the density term.

[0054] S3. Perform hierarchical differential search processing on the legitimate primary population and the legitimate auxiliary population according to the mating fitness value to obtain the offspring population.

[0055] In this embodiment of the invention, step S3 includes the following sub-steps: S31. Based on the binary tournament mechanism, individuals are extracted from the legitimate main population to obtain a first extracted parent main population. The first extracted parent main population is sorted in ascending order by the mating fitness and split into a superior subset and a inferior subset. Diversity difference is performed on the superior subset and convergence difference is performed on the inferior subset to obtain the difference result. Boundary pruning and polynomial mutation processing are performed on the difference result to obtain the main population difference subset.

[0056] Specifically, a mating sample of at least the population size is first drawn using a binary tournament. This sample is then divided into two subsets, one relatively superior and one relatively inferior, in ascending order of mating fitness. The superior subset undergoes differential sampling towards greater diversity. (The three positions are distinct within the superior subset); the inferior subset performs a difference quantization toward convergence. ,in It is a randomly selected superior individual rather than the globally optimal individual; that is, it "moves towards a randomly better direction" rather than the standard DE / best / 1. The scaling factor is taken independently and with equal probability in each generated row. This approach preserves medium and large strides, reducing the search's sensitivity to a single scale. Site selection employs two mutually exclusive modes: the grouping mode applies difference and polynomial mutation only once per random channel (all layers within that channel), preserving the inter-layer structure within the channel; the integer vector mode independently extracts the coordinate-wise acceptance rate for each individual. Determine which coordinates accept differencing, and in The coordinate probabilities are subjected to multinomial mutation. The selection probabilities of the two modes are: ; As the evaluation budget progresses It dropped rapidly from 1 to about 0.15 ( ) and 0.013 ( This reflects a search rhythm arrangement of "early focus on single-channel structural perturbations, later focus on cross-channel coordination." Differential results are first trimmed to... Then apply a distribution index to the selected coordinates. The bounded polynomial mutation (the larger the exponent, the more concentrated the perturbation is near the parent, while retaining a small probability of a longer step size) is performed, followed by pruning.

[0057] Polynomial mutation is a bounded real-number mutation operator widely used in NSGA-II / PlatEMO. Given the current value of the selected coordinates. Upper and lower bounds of a single variable With distribution index First, calculate the relative distance from the point to the two boundaries. , ; draw again Obtain the offset ratio: ; final The formula contains The terms provide the boundedness that "the closer to the boundary, the smaller the offset in the direction of that boundary is compressed"; The larger The more concentrated the value is at 0, the closer the perturbation is to the parent generation. This is a conservative setting, beneficial for refining near the established structure while preserving a low probability of longer jumps. Unlike differential mutation, which depends on population distribution, multinomial mutation is an independent random perturbation unrelated to the population, used to maintain minimum coordinate-level diversity and prevent premature convergence and locking of individual dimensions; it prunes both before and after mutation to... .

[0058] S32. Perform the same processing logic as in step S31 on the legal auxiliary population to obtain the auxiliary population difference numerator; S33. Random individuals are randomly extracted from the legal main population to obtain a second extracted parent main population; the average channel film thickness of each channel in the second extracted parent main population is adjusted according to the normalized average transmittance constraint residual corresponding to the second extracted parent main population, and a perturbation is applied to the film thickness of a random number of channels in the adjusted second extracted parent main population to obtain a thickness adjustment subset.

[0059] Specifically, the second parent population is uniformly drawn from the legitimate primary population, and its normalized average transmittance constrained residual is directly read. .like (Insufficient or near-critical transmittance;) Corresponding to the threshold The lower transmittance is exactly 0.51, which is the absolute margin of 0.01 reserved at the feasible boundary, under the following conditions: ; The average total channel thickness of the second-generation parent population was reduced by approximately [percentage missing]. – To increase permeability by reducing thickness; if (If there is already a margin), then proceed according to the following conditions: ; Upward adjustment – To restore the expressive power of interference modulation. Thinning is generally beneficial for improving transmittance, but it is not strictly monotonic for multilayer interference structures; therefore, this operator is a physically heuristic adjustment rather than a strictly constrained projection. After the target total thickness is trimmed to the reachable range, it is adjusted channel by channel to the same target total thickness, and then adjusted from approximately... ( In each channel, the amplitude is uniformly added to the odd-numbered layers and uniformly subtracted from the even-numbered layers. ( , The disturbance (which decreases linearly with progress) is introduced to introduce sequence structure differences while keeping the total thickness constant. The magnitude of the disturbance narrows with progress to balance early structural exploration and later refinement.

[0060] S34. Randomly extract individuals from the legitimate main population to obtain a baseline parent population; uniformly and randomly extract individuals from the legitimate main population and the legitimate auxiliary population to obtain a donor population; construct a cost matrix between the baseline parent population and the donor population using the Euclidean distance of the channel thickness vector; repeatedly select the minimum cost pair that has not yet been occupied in the cost matrix based on a greedy algorithm; map and align the channels of the baseline parent population with the channels of the donor population one by one; select a random number of channels in the baseline parent population to perform whole channel replacement or layer thickness convex combination interpolation and recombination and apply perturbation to obtain an aligned recombination subset.

[0061] Specifically, since the three optimization objectives are approximately invariant to the channel arrangement, different channel permutations may represent the design of the same optical effect. The operator first uniformly extracts a baseline parent population from the legitimate main population, and then uniformly and randomly extracts individuals from the legitimate main population and the legitimate auxiliary population to obtain the donor population. A cost matrix is ​​constructed using the Euclidean distance of the channel thickness vector, and a greedy algorithm is used to repeatedly select the minimum cost pair that has not yet been occupied to obtain a one-to-one alignment between the parent and donor channels (not a globally optimal assignment, but only to alleviate, not strictly eliminate, the permutation symmetry).

[0062] The baseline parent population and the donor population were restored to their respective values. Channel-layer matrix , ,in Indicates the first Each channel starts from the incident side Layer thickness vector. The meanings of the symbols and formulas are as follows.

[0063] Cost matrix aligned with greedy algorithm: Construct the cost matrix using the Euclidean distance between the parent channel and the donor channel. .

[0064] ; Right now The smaller the value, the higher the paternity. Channel and Donor The more similar the channel structures, the more greedy the alignment process repeatedly selects the smallest value from the unoccupied rows and columns. Lock the line With column Until a one-to-one mapping is obtained. (donor channel) Corresponding to the parent channel This is a heuristic that gradually finds the local optimum, not a global optimum assignment; therefore, it can only alleviate, not strictly eliminate, the problem. Permutation symmetry. Below... All refer to those that have been pressed Aligned donor channels.

[0065] Whole row replacement and convex combination: Randomly select from all channels. Each channel participates in the reorganization, and each selected channel... by The probability of replacing the entire line with the aligned donor channel (Preserve the interlayer coupling within the channel), otherwise perform independent two-point convex combination for each channel: ; in The mixing coefficient for this channel (each participating channel is extracted independently), since And the sum of the two weights is The resulting thickness of each layer falls between the corresponding layers of the parent and the donor. More like the father generation More like a donor) Round it to the nearest integer thickness.

[0066] Odd-even layer reverse perturbation: then with A structural perturbation is added to a randomly selected channel, with the same amplitude added to odd-numbered layers and subtracted from even-numbered layers of that channel: ; in For standard normal random numbers, (Unit: nm) represents the progress of the evaluation. The perturbation scale of linear narrowing (early approximately) Later Approx. The inverse form of odd-plus-even-minus alters the sequence structure while approximately maintaining the total channel thickness, similar to the perturbation in the transmittance operator (whose scale is...). They share the same design philosophy; the scope narrows as progress is made, taking into account both early structural exploration and later refinement.

[0067] S35. Merge the main population difference subset, the auxiliary population difference subset, the thickness adjustment subset, and the alignment recombination subset to obtain the candidate offspring population; Specifically, after each iteration, the actual evaluation quota is determined based on the remaining budget. Then according to the oversampling factor (Can be set to 2) to obtain the candidate size Of which approximately (Can be set to 0.30) share The remaining steps are assigned to two problem-specific operators (i.e., steps S33 and S34). Distribute to two differential search paths (i.e., steps S31 and S32); default full quota ( When ), the request quantities corresponding to steps S31-S34 are approximately respectively. .

[0068] S36. The candidate offspring population is modified by the manufacturing feasibility constraints and candidate deduplication is performed to obtain the offspring population.

[0069] Specifically, after the four candidate streams have been made feasible, stable deduplication is performed only within the current generation's candidate pool based on the complete decision vector (the order of first appearance is retained, and duplicates are discarded without completion). If a large number of candidates are mapped to the same integer design after repair (e.g., multiple parent generations perturb near the same channel structure), the actual number of unique candidates may be significantly less than the requested quota, thereby compressing the number of subsequent real evaluations and the amount of surrogate training samples. This deduplication only applies to the current generation's candidate pool and is not compared with historically evaluated designs. Therefore, the same design may still be repeatedly evaluated in different generations.

[0070] S4. Based on the ridge regression surrogate model, the offspring population is screened, and optical simulation is performed on each genetic individual in the screened offspring population using the transfer matrix method to obtain the spectral modulation matrix of the corresponding filter array; an optimization target is established, and the nominal performance value corresponding to the optimization target is calculated based on the spectral modulation matrix.

[0071] Based on performance constraints, each genetic individual in the offspring population is verified and labeled, and the consumed evaluation budget is updated to obtain a constraint feasibility label.

[0072] In this embodiment of the invention, because the transfer matrix method is costly, a low-cost multi-output linear ridge regression model is used to perform a preliminary screening of the offspring population before evaluation. If the number of unique candidates does not exceed the evaluation quota... All candidates proceed directly to the actual evaluation; otherwise, the current two populations are used ( The unique samples after deduplication are used as the training set. If the number of unique samples is less than [a certain number], then [the training set is not specified]. ( (For the target number), the sample size is too small to fit the multi-output model, degenerating into randomly selecting from the candidates. One. Under the normal path, decision variables are set according to fixed upper and lower bounds. – Normalization: The objective and each constraint column are normalized according to the minimum and range of the training set columns (zero range is replaced with 1 to avoid division by zero), and then concatenated into a joint output. After adding the intercept column, solve for the ridge regression coefficients:

[0073] Regularization coefficient Scaling by average feature energy is motivated by the fact that the number of training samples is typically small. When (underdetermined) numerical conditions are improved and solvability is guaranteed, the proportionality constant is used to improve the numerical conditions. These are empirically set parameters, without cross-validation or parameter tuning. After denormalizing the predicted values, each constraint column is relaxed according to the first boundary. (here This involves relaxing constraints column by column (different from the application to the total violation in subsequent environment selection). Based on this, a constrained non-dominated sort is performed, selecting the top columns in the order of "predicted frontier number ascending, predicted total violation number ascending, and random number breaking the tie." The surrogate model only determines which candidates are worth spending real evaluation resources on; the predicted values ​​are not written into the population to replace the real evaluation results.

[0074] Selected candidates undergo realistic Transmission Matrix Method (TMM) optical simulations after passing through a unified evaluation interface. This calculates three objectives and the currently enabled performance constraints, incrementing the FE by 1 for each evaluated design. Under the Pareto dominance strategy (nominal mode), each design corresponds to one array-level spectral calculation (covering all parameters). Propagation of membrane stacks in individual channels); robustness assessment strategy (robust model) ( The next design requires internal considerations. Subarray-level spectral calculations are performed, but regardless of the internal computational load, FE is only incremented by 1.

[0075] Step S4 includes the following sub-steps: S41. For each genetic individual in the offspring population, forward simulation is performed using the transfer matrix method to obtain the spectral transmittance curves of each filter channel in the full band. Multiple spectral transmittance curves are spliced ​​together to obtain the spectral modulation matrix. S42. Calculate the nominal performance value corresponding to the optimization objective based on the spectral modulation matrix; wherein, the optimization objective includes an information volume measurement objective, a spectral response structure independence objective, and an inverse problem benignity objective: The information volume measurement objective is: based on the Bayesian D-optimality criterion, to measure the total amount of information captured from the scene by the filter array represented by the spectral modulation matrix, and to drive the optimization of the light flux, channel independence and band targeting in the corresponding dimension of the spectral modulation matrix; The objective of spectral response structure independence is to measure the non-redundancy of the encoded information of each filter channel of the spectral modulation matrix by decentering the spectral response curves of each filter channel of the spectral modulation matrix and introducing a wavelength offset. The benignity objective of the inverse problem is: the degree of amplification of the noise in the inversion and reconstruction after the observation signal is encoded by the spectral modulation matrix, which is quantified by the conditions of the spectral modulation matrix; S43. Verify whether each genetic individual in the offspring population satisfies the performance constraint, mark genetic individuals that satisfy the performance constraint as feasible individuals, and mark genetic individuals that do not satisfy the performance constraint as infeasible individuals, thus obtaining constraint feasibility labels; wherein, the ranking priority of any feasible individual is higher than that of all infeasible individuals, and the performance constraint is to make the offspring population satisfy the following conditions: The average transmittance of all filter channels in the offspring population is not lower than a preset transmittance threshold. The standard deviation of the average transmittance among the filter channels in the offspring population is less than a preset standard deviation threshold. The spectral resolution of all filter channels in the offspring population is not lower than the preset spectral resolution.

[0076] In embodiments of the present invention, such as Figure 2 As shown, Figure 2 This is a schematic diagram of the spectral transmittance curve of the filter array obtained by forward simulation using the transfer matrix method according to an embodiment of the present invention. Figure 2 With 6 Taking the layout of 6 as an example, a set of filter transmittance curves obtained through TMM forward simulation are given. Each curve in the figure corresponds to the spectral transmittance of a filter channel in the array within the visible light band as a function of wavelength: the horizontal axis represents wavelength, and the vertical axis represents transmittance. It can be seen that the transmittance curves of different channels exhibit significant differences in peak position, bandwidth, and modulation pattern. It is this difference that constitutes the multi-channel encoding capability of the incident spectrum, and is also the core object measured by the various evaluation indicators of this invention.

[0077] S5. Based on the nominal performance value and the constraint feasibility marker, perform environmental selection on the offspring population to obtain the next generation main population and the next generation auxiliary population; wherein, when selecting the next generation main population, the selection is based on a preset boundary from the legal main population and the offspring population; when selecting the next generation auxiliary population, the selection is based on a dynamic boundary from the legal auxiliary population, the offspring population and the next generation main population.

[0078] In this embodiment of the invention, after a real evaluation, environmental selection based on an additive logarithmic index determines which individuals in the offspring population will enter the next generation. First, the merged pool of the current population (legitimate main population, legitimate auxiliary population) and the new offspring (offspring population) is stabilized and deduplicated according to the decision vector: if the number of unique individuals does not exceed... If all individuals are retained, the strength-density fraction included in the constraint is returned; otherwise, each target is shifted to a positive value. ( (Avoid zero in logarithms or ratios) Define the reference set (if empty, press Enter) ascending order, first order (A set of alternative references).

[0079] For each design in the population Take all the remaining designs As "challengers", we construct comparison values ​​one by one. Environmental indicators are among the most unfavorable. (The higher the value, the better). It's important to specify the direction of the comparison: when designing... When it belongs to the reference set, The value is determined by the challenger The violation amount determines the branch, if Based solely on the logarithmic ratio of each objective Comparison (the challenger's variation on any objective is taken as the largest logarithmic ratio, representing the worst-case deterioration; if the challenger's variation on all objectives is not worse, the smallest logarithmic ratio is taken, representing the strength of dominance); if Then, the logarithm of the larger of the target multiplier and the constraint multiplier is taken. That is, whether or not constraints need to be compared simultaneously depends on the degree of violation by the challenger, not on the design. The degree of violation itself. When designing When it does not belong to the reference set, a formula based on relative violations is used instead, mainly composed of... The ratio of the target span and the violation amount of the joint pool, along with the global offset term determined by the target span, make the overall performance of this type of design significantly lower, and it will only be retained when the size of the reference set is insufficient to fill the population capacity.

[0080] According to the first-level set Decision on subsequent handling: If Directly in all designs (including but not limited to) (in) Take first in descending order One, the insufficient part is used Design improvements; if Only in Internally, a reference vector partitioning and truncation process is performed. After normalizing both the forward-moving target and the reference vector, they are assigned to the nearest partition based on their maximum inner product. This process is repeated, removing the partition with the smallest index from the partition with the most current members, until only a few remain. This truncation balances convergence (prioritizing designs that ensure positive metrics, i.e., "not being completely defeated by any challenger") and diversity (avoiding overcrowding in the same partition). Strict boundaries are used for the main population. The auxiliary population uses dynamic boundaries calculated based on an internal dynamic clock, and both reuse the same environmental selection process.

[0081] The main population and auxiliary populations independently complete the above-mentioned environmental selection: the next generation main population only merges the legitimate main population and the offspring population, and uses strict boundary screening to ensure that the designs entering the next generation reference set truly meet the performance constraints; the next generation auxiliary population merges the legitimate auxiliary population, the next generation main population and the offspring population, and uses dynamic boundary screening, thus it can absorb the high-quality feasible regions just generated by the main population.

[0082] S6. Determine whether the updated consumed evaluation budget has reached or exceeded the preset evaluation budget limit: If so, extract all non-dominated solutions from the current next-generation master population to obtain the Pareto optimal filter array design solution set, calculate the peak signal-to-noise ratio corresponding to each candidate design in the Pareto optimal filter array design solution set, and output the candidate design with the largest peak signal-to-noise ratio as the final design scheme of the hyperspectral computational imaging encoder. If not, then the next generation main population is taken as the new legitimate main population, and the next generation auxiliary population is taken as the new legitimate auxiliary population, and the process returns to step S2 to continue iterative optimization.

[0083] In this embodiment of the invention, after the evolutionary cycle terminates, a single final deployment design needs to be selected from the returned population. Since the three evolutionary objectives (information volume, spectral structure independence, and inverse problem benignity) are all indirect proxies of reconstruction quality, the final selection uses a more application-oriented metric: each candidate filter array is tested in an actual hyperspectral image reconstruction task, and the one with the highest reconstruction quality is selected. This step is a one-time post-processing step after evolution, which is not included in the FE budget and does not affect the population.

[0084] First, screen out the next generation of main populations that meet all performance constraints. The feasible design of the filter array is obtained by performing a non-dominated sorting of its objectives, and then obtaining the Pareto optimal filter array design solution set. Only in Reconstruction assessments should be conducted to avoid wasting reconstruction calculations on dominated or infeasible designs.

[0085] Reconstruction and scoring: For each candidate design , by its aisle, transmittance response matrix at wavelength Constructing a perception (coding) matrix for a given reference hyperspectral image First, obtain the measurement through forward imaging. Then use the selected reconstruction algorithm Inverse solution to reconstruct results The fidelity of the reconstruction is measured by peak signal-to-noise ratio (PSNR). ; in, The maximum possible value of an image pixel (e.g., normalized to) Or 8-bit ), The total number of pixels-bands included in the statistics. PSNR represents the mean squared error between the reconstructed image and the ground truth image. A higher PSNR indicates that the reconstruction is closer to the ground truth.

[0086] Optimal Individual Determination: The candidate with the highest PSNR is selected as the final design.

[0087] If the evaluation is performed on multiple images or the entire test set, the mean value of the PSNR of each design should be used as the selection criterion (and if necessary, the evolutionary objective or secondary reconstruction metric can be compared when comparing them side by side).

[0088] Compared with existing technologies, this invention constructs a multi-dimensional evaluation system of optimization objectives based on the inherent physical characteristics of the spectral modulation matrix. This system consists of an information volume measurement objective, a spectral response structure independence objective, and an inverse problem benignity objective. The nominal performance value is used as the basis for optimization. Performance evaluation and optimization can be completed without relying on specific spectral reconstruction algorithms and training datasets. The resulting filter array scheme is not affected by changes in the backend reconstruction algorithm and has cross-algorithm versatility. At the same time, it avoids the computational power consumption caused by a large number of reconstruction iteration calculations, and significantly improves design efficiency.

[0089] This invention directly uses the thickness of each thin film layer in the filter thickness matrix as a design variable, and generates a new film system structure through an evolutionary algorithm, rather than combining and screening from a preset candidate library. It can explore non-classical structural forms outside of traditional classic film systems, break the constraints of the candidate library on the design space, and obtain filter array schemes with stronger encoding capabilities and better spectral modulation characteristics.

[0090] This invention uses three mutually balancing performance indicators as parallel optimization objectives and obtains a Pareto-optimal filter array design solution set composed of non-dominated solutions through multi-objective evolutionary search. By postponing the engineering trade-off step to the end of the optimization process, the final solution can be selected from the leading solution set according to the performance emphasis of different application scenarios, thus avoiding performance bias caused by subjective weighted summation.

[0091] This invention employs a hybrid initialization strategy combining constrained random initialization, quarter-wavelength optical thickness initialization, and Fabry-Perot cavity initialization to balance the structural rationality and morphological diversity of the initial population. It utilizes crossover operations with the complete thickness vector corresponding to a single filter as the crossover unit, combined with sparse adaptive Gaussian mutation and random replacement mutation, to improve evolutionary efficiency while preserving the effective interlayer coordination structure. Furthermore, it achieves adaptive environment selection by integrating a comprehensive performance index based on diversity evaluation of the target space sub-region division and convergence and robustness values ​​calculated using a scalarization function. This dynamically adjusts the exploration direction according to the population state, balancing convergence speed and population diversity.

[0092] At the constraint and control level, the initial candidate population and offspring candidate population are modified in real time by manufacturing feasibility constraints, and legal initial population and legal offspring population are output throughout the process. Combined with the dominance mechanism of performance soft constraints and constraint feasibility markers, it ensures that all candidate solutions are within the process feasible domain and meet the basic performance requirements such as transmittance, uniformity, and spectral resolution, thus avoiding the generation of invalid solutions with no engineering application value.

[0093] This invention employs strict constraint screening on the main population, focusing on multi-objective convergence within the feasible region to ensure the constraint compliance and Pareto front quality of the final solution. The auxiliary population uses dynamic relaxation constraints, retaining individuals that are temporarily infeasible but possess excellent target performance and exploration value, effectively expanding the search space, maintaining population diversity, and helping the algorithm traverse infeasible regions and escape local optima. Both populations jointly generate candidate solutions, and the offspring of truly evaluated solutions participate in the dual-population update simultaneously. Simultaneously, the updated high-quality solutions from the main population are directionally injected into the auxiliary population, providing feasibility guidance for the exploration process and preventing the search from deviating from the feasible direction, thereby achieving multi-directional collaboration and improved efficiency. The overall architecture achieves a synergistic balance between global exploration, constraint satisfaction, and multi-objective convergence. While ensuring the design meets manufacturing and performance constraints, it effectively improves the algorithm's global optimization capability and convergence efficiency, resulting in a hyperspectral computational imaging encoder design with superior reconstruction performance.

[0094] Example 2 This invention also provides a design system for a hyperspectral computational imaging encoder, please refer to... Figure 3 , Figure 3 This is a schematic diagram of the design system 200 for a hyperspectral computational imaging encoder provided in an embodiment of the present invention, which includes: The initialization module 201 is used to construct an initial primary candidate population and an initial auxiliary candidate population based on a hybrid initialization strategy, using the thickness of each thin film layer of the filter array as design variables. The initial primary candidate population and the initial auxiliary candidate population are then modified by manufacturing feasibility constraints to obtain a legal primary population and a legal auxiliary population. The initial primary candidate population and the initial secondary candidate population each include multiple genetic individuals, with each genetic individual corresponding to a group. M×N L The filter thickness matrix, M The number of filters in the filter array. N L The number of thin film layers in a single filter in the filter array, wherein each filter comprises sequentially arranged... N L One gene unit, N L Each of the aforementioned gene units corresponds to a filter layer from the first layer to the second layer. N L The thickness of the thin film layer; The mating fitness calculation module 202 is used to calculate the budget progress based on the current consumed evaluation budget and the preset evaluation budget limit, and to perform overall stagnation detection on the legitimate main population to obtain the stagnation detection result; select the corresponding strategy based on the budget progress and the stagnation detection result, and calculate the mating fitness values ​​corresponding to the legitimate main population and the legitimate auxiliary population; The hierarchical differential module 203 is used to perform hierarchical differential search processing on the legitimate primary population and the legitimate auxiliary population according to the mating fitness value to obtain the offspring population; The optimization target module 204 is used to screen the offspring population based on the ridge regression surrogate model, and to perform optical simulation on each genetic individual in the screened offspring population using the transfer matrix method to obtain the spectral modulation matrix of the corresponding filter array; to establish an optimization target, and to calculate the nominal performance value corresponding to the optimization target based on the spectral modulation matrix; Based on performance constraints, each genetic individual in the offspring population is verified and labeled, and the consumed evaluation budget is updated to obtain a constraint feasibility label. The screening module 205 is used to perform environmental selection on the offspring population based on the nominal performance value and the constraint feasibility marker, and to screen out the next generation main population and the next generation auxiliary population; wherein, when screening the next generation main population, the screening is based on a preset boundary from the legal main population and the offspring population; when screening the next generation auxiliary population, the screening is based on a dynamic boundary from the legal auxiliary population, the offspring population and the next generation main population; The judgment module 206 is used to determine whether the updated consumed evaluation budget has reached or exceeded the preset evaluation budget limit: If so, extract all non-dominated solutions from the current next-generation master population to obtain the Pareto optimal filter array design solution set, calculate the peak signal-to-noise ratio corresponding to each candidate design in the Pareto optimal filter array design solution set, and output the candidate design with the largest peak signal-to-noise ratio as the final design scheme of the hyperspectral computational imaging encoder. If not, the next generation primary population is taken as the new legitimate primary population, and the next generation secondary population is taken as the new legitimate secondary population. The process then returns to the mating fitness calculation module 202 to continue iterative optimization.

[0095] The hyperspectral computational imaging encoder design system 200 can implement the steps in the hyperspectral computational imaging encoder design method as described in the above embodiments, and can achieve the same technical effect. Refer to the description in the above embodiments, which will not be repeated here.

[0096] Example 3 This invention also provides a computer device, please refer to... Figure 4 , Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. The computer device 300 includes: a memory 302, a processor 301, and a design program for a hyperspectral computational imaging encoder stored in the memory 302 and capable of running on the processor 301.

[0097] The processor 301 calls the design program for the hyperspectral computational imaging encoder stored in the memory 302, and executes the steps in the design method of the hyperspectral computational imaging encoder provided in this embodiment of the invention. Please refer to... Figure 1 Specifically, it includes the following steps: S1. Using the thickness of each thin film layer of the filter array as the design variable, construct an initial primary candidate population and an initial auxiliary candidate population based on a hybrid initialization strategy. Modify the initial primary candidate population and the initial auxiliary candidate population respectively through manufacturing feasibility constraints to obtain a legal primary population and a legal auxiliary population. The initial primary candidate population and the initial secondary candidate population each include multiple genetic individuals, with each genetic individual corresponding to a group. M×N L The filter thickness matrix, M The number of filters in the filter array. N L The number of thin film layers in a single filter in the filter array, wherein each filter comprises sequentially arranged... N L One gene unit, N L Each of the aforementioned gene units corresponds to a filter layer from the first layer to the second layer. NL The thickness of the thin film layer; S2. Calculate the budget progress based on the current consumed evaluation budget and the preset evaluation budget limit, and perform an overall stagnation detection on the legitimate main population to obtain the stagnation detection result; select the corresponding strategy based on the budget progress and the stagnation detection result, and calculate the mating fitness values ​​corresponding to the legitimate main population and the legitimate auxiliary population. S3. Perform hierarchical differential search processing on the legitimate primary population and the legitimate auxiliary population according to the mating fitness value to obtain the offspring population; S4. The offspring population is screened based on the ridge regression surrogate model, and optical simulation is performed on each genetic individual in the screened offspring population using the transfer matrix method to obtain the spectral modulation matrix of the corresponding filter array; an optimization target is established, and the nominal performance value corresponding to the optimization target is calculated based on the spectral modulation matrix. Based on performance constraints, each genetic individual in the offspring population is verified and labeled, and the consumed evaluation budget is updated to obtain a constraint feasibility label. S5. Based on the nominal performance value and the constraint feasibility marker, perform environmental selection on the offspring population to obtain the next generation main population and the next generation auxiliary population; wherein, when selecting the next generation main population, the selection is based on a preset boundary from the legal main population and the offspring population; when selecting the next generation auxiliary population, the selection is based on a dynamic boundary from the legal auxiliary population, the offspring population and the next generation main population; S6. Determine whether the updated consumed evaluation budget has reached or exceeded the preset evaluation budget limit: If so, extract all non-dominated solutions from the current next-generation master population to obtain the Pareto optimal filter array design solution set, calculate the peak signal-to-noise ratio corresponding to each candidate design in the Pareto optimal filter array design solution set, and output the candidate design with the largest peak signal-to-noise ratio as the final design scheme of the hyperspectral computational imaging encoder. If not, then the next generation main population is taken as the new legitimate main population, and the next generation auxiliary population is taken as the new legitimate auxiliary population, and the process returns to step S2 to continue iterative optimization.

[0098] The computer device 300 provided in this embodiment of the invention can implement the steps in the method as described in the above embodiments and can achieve the same technical effect. Refer to the description in the above embodiments, which will not be repeated here.

[0099] Example 4 This invention also provides a computer-readable storage medium storing a program. When the program is executed by a processor, it implements the various processes and steps in the method provided in this invention and achieves the same technical effects. To avoid repetition, it will not be described again here.

[0100] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by hardware related to computer programs or instructions. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0101] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0102] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0103] The embodiments of the present invention have been described above with reference to the accompanying drawings. The disclosed embodiments are merely preferred embodiments of the present invention. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many equivalent changes in form under the guidance of the present invention without departing from the spirit and scope of the claims. All such changes are within the protection scope of the present invention.

Claims

1. A design method for a hyperspectral computational imaging encoder, characterized in that, The design method includes the following steps: S1. Using the thickness of each thin film layer of the filter array as the design variable, construct an initial primary candidate population and an initial auxiliary candidate population based on a hybrid initialization strategy. Modify the initial primary candidate population and the initial auxiliary candidate population respectively through manufacturing feasibility constraints to obtain a legal primary population and a legal auxiliary population. The initial primary candidate population and the initial secondary candidate population each include multiple genetic individuals, with each genetic individual corresponding to a group. M×N L The filter thickness matrix, M The number of filters in the filter array. N L The number of thin film layers in a single filter in the filter array, wherein each filter comprises sequentially arranged... N L One gene unit, N L Each of the aforementioned gene units corresponds to a filter layer from the first layer to the second layer. N L The thickness of the thin film layer; S2. Calculate the budget progress based on the current consumed evaluation budget and the preset evaluation budget limit, and perform an overall stagnation detection on the legitimate main population to obtain the stagnation detection result; select the corresponding strategy based on the budget progress and the stagnation detection result, and calculate the mating fitness values ​​corresponding to the legitimate main population and the legitimate auxiliary population. S3. Perform hierarchical differential search processing on the legitimate primary population and the legitimate auxiliary population according to the mating fitness value to obtain the offspring population; S4. The offspring population is screened based on the ridge regression surrogate model, and optical simulation is performed on each genetic individual in the screened offspring population using the transfer matrix method to obtain the spectral modulation matrix of the corresponding filter array; an optimization target is established, and the nominal performance value corresponding to the optimization target is calculated based on the spectral modulation matrix. Based on performance constraints, each genetic individual in the offspring population is verified and labeled, and the consumed evaluation budget is updated to obtain a constraint feasibility label. S5. Based on the nominal performance value and the constraint feasibility marker, perform environmental selection on the offspring population to obtain the next generation main population and the next generation auxiliary population; wherein, when selecting the next generation main population, the selection is based on a preset boundary from the legal main population and the offspring population; when selecting the next generation auxiliary population, the selection is based on a dynamic boundary from the legal auxiliary population, the offspring population and the next generation main population; S6. Determine whether the updated consumed evaluation budget has reached or exceeded the preset evaluation budget limit: If so, extract all non-dominated solutions from the current next-generation master population to obtain the Pareto optimal filter array design solution set, calculate the peak signal-to-noise ratio corresponding to each candidate design in the Pareto optimal filter array design solution set, and output the candidate design with the largest peak signal-to-noise ratio as the final design scheme of the hyperspectral computational imaging encoder. If not, then the next generation main population is taken as the new legitimate main population, and the next generation auxiliary population is taken as the new legitimate auxiliary population, and the process returns to step S2 to continue iterative optimization.

2. The design method of the hyperspectral computational imaging encoder as described in claim 1, characterized in that, In step S1, the hybrid initialization strategy is as follows: the film thickness of the filter for the genetic individual is generated through constrained random initialization, quarter-wavelength optical thickness initialization, and Fabry-Perot cavity initialization. The constraint is randomly initialized as follows: the total thickness value of a single filter is randomly generated within a preset total thickness range, and the total thickness value is proportionally distributed to each layer of the filter film using random weights. The quarter-wavelength optical thickness is initialized by randomly selecting the center wavelength, calculating the thickness of each thin film layer of the filter based on the quarter-wavelength interference condition, and applying multiplicative Gaussian noise perturbation. The Fabry-Perot cavity is initialized by amplifying the film thickness value corresponding to the intermediate layer of the filter based on the quarter-wavelength basic film system structure and applying multiplicative noise perturbation.

3. The design method of the hyperspectral computational imaging encoder as described in claim 1, characterized in that, The manufacturing feasibility constraints include material system constraints, structural size constraints, single-layer thickness constraints, and total thickness constraints: The material system constraint is that the filter adopts a preset high and low refractive index material system and a fixed alternating stacking order; The structural size constraint is that the number of thin film layers of the filter and the total number of channels in the filter array are both preset fixed values; The single-layer thickness constraint means that the film thickness of each layer of the filter must be within the preset upper and lower limits of the single-layer thickness. The total thickness constraint means that the total thickness of the filter must be within a preset total thickness target range.

4. The design method of the hyperspectral computational imaging encoder as described in claim 1, characterized in that, Step S2 also includes the following sub-steps: S21. Establish an internal generation clock, calculate the dynamic boundary based on the internal generation clock, and update the boundary of the legitimate auxiliary population based on the dynamic boundary; wherein, the internal generation clock is dynamically adjusted according to the number of population iterations; S22. Perform overall stagnation detection on the legitimate main population to obtain stagnation detection results; calculate the budget progress based on the current consumed evaluation budget and the preset evaluation budget upper limit; S23. When the stagnation detection result is stagnation or the budget progress exceeds the preset budget progress threshold, the mating fitness values ​​corresponding to the legitimate primary population and the legitimate auxiliary population are calculated based on the robust evaluation strategy. When the stagnation detection result is no stagnation or the budget progress does not exceed the preset budget progress threshold, the mating fitness values ​​corresponding to the legitimate primary population and the legitimate secondary population are calculated using the Pareto dominance strategy.

5. The design method of the hyperspectral computational imaging encoder as described in claim 4, characterized in that, The robustness evaluation strategy is as follows: The legitimate primary population and the legitimate secondary population are deduplicated and merged to obtain a mixed population; The convergence of each genetic individual in the mixed population is quantified using an achievement scalarization function, resulting in multiple convergence values. The achievement scalarization function satisfies the following condition: ； in, Indicates genetic individual x In the The target value on each target, This indicates that all current genetic individuals are in the [number]th [year]. The optimal value for each objective. Indicates genetic individual x In the r A preference weight vector on each objective The achievement scalarization function is represented by the convergence value obtained by normalizing the result of the achievement scalarization function. ; A manufacturing error perturbation is applied to the genetic individuals of the mixed population to obtain a perturbation target value. The sum of the deviations of the genetic individuals of the mixed population from the perturbation target value is calculated as a robustness value. The robustness value satisfies the following condition: ； in, This represents the robustness value. Indicates genetic individual x In the r Deviation from the target Indicates genetic individual x After the disturbance The perturbation target value on each target; A comprehensive performance index is calculated based on the robustness value and the convergence value, and this comprehensive performance index is used as the mating fitness value. The comprehensive performance index satisfies the following conditions: ； ； in, Indicates adaptive weights, This represents the median of the achievement scalarization function results for all genetically derived individuals in the mixed population. and This represents the historical maximum and minimum values ​​of the achievement scalarization function during the iteration process. Indicates genetic individual x Corresponding comprehensive performance indicators; The Pareto domination strategy is as follows: domination relationships are determined for all genetic individuals in the mixed population pairwise, and a constraint priority rule is introduced in the domination determination process; wherein, the constraint priority rule is: when comparing two genetic individuals, the genetic individual with a smaller total constraint violation has a higher priority. The number of other individuals dominated by each genetically derived individual is counted to obtain the intensity value; The density score is obtained by calculating the neighbor distances of each genetic individual in the target space corresponding to the optimization objective to the neighboring individuals. The comprehensive score of each genetic individual is calculated based on the intensity value and the density score, and the comprehensive score is used as the mating fitness value.

6. The design method of the hyperspectral computational imaging encoder as described in claim 1, characterized in that, Step S4 includes the following sub-steps: S41. For each genetic individual in the offspring population, forward simulation is performed using the transfer matrix method to obtain the spectral transmittance curves of each filter channel in the full band. Multiple spectral transmittance curves are spliced ​​together to obtain the spectral modulation matrix. S42. Calculate the nominal performance value corresponding to the optimization objective based on the spectral modulation matrix; wherein, the optimization objective includes an information volume measurement objective, a spectral response structure independence objective, and an inverse problem benignity objective: The information volume measurement objective is: based on the Bayesian D-optimality criterion, to measure the total amount of information captured from the scene by the filter array represented by the spectral modulation matrix, and to drive the optimization of the light flux, channel independence and band targeting in the corresponding dimension of the spectral modulation matrix; The objective of spectral response structure independence is to measure the non-redundancy of the encoded information of each filter channel of the spectral modulation matrix by decentering the spectral response curves of each filter channel of the spectral modulation matrix and introducing a wavelength offset. The benignity objective of the inverse problem is: the degree of amplification of the noise in the inversion and reconstruction after the observation signal is encoded by the spectral modulation matrix, which is quantified by the conditions of the spectral modulation matrix; S43. Verify whether each genetic individual in the offspring population satisfies the performance constraint, mark genetic individuals that satisfy the performance constraint as feasible individuals, and mark genetic individuals that do not satisfy the performance constraint as infeasible individuals, thus obtaining constraint feasibility labels; wherein, the ranking priority of any feasible individual is higher than that of all infeasible individuals, and the performance constraint is to make the offspring population satisfy the following conditions: The average transmittance of all filter channels in the offspring population is not lower than a preset transmittance threshold. The standard deviation of the average transmittance among the filter channels in the offspring population is less than a preset standard deviation threshold. The spectral resolution of all filter channels in the offspring population is not lower than the preset spectral resolution.

7. The design method of the hyperspectral computational imaging encoder as described in claim 1, characterized in that, Step S3 includes the following sub-steps: S31. Based on the binary tournament mechanism, individuals are extracted from the legitimate main population to obtain the first extracted parent main population. The first extracted parent main population is sorted in ascending order by the mating fitness and split into a superior subset and a inferior subset. Diversity difference is performed on the superior subset and convergence difference is performed on the inferior subset to obtain the difference result. The difference results are subjected to boundary pruning and polynomial mutation to obtain the differential set of the principal population. S32. Perform the same processing logic as in step S31 on the legal auxiliary population to obtain the auxiliary population difference numerator; S33. Random individuals are randomly extracted from the legal main population to obtain a second extracted parent main population; the average channel film thickness of each channel in the second extracted parent main population is adjusted according to the normalized average transmittance constraint residual corresponding to the second extracted parent main population, and a perturbation is applied to the film thickness of a random number of channels in the adjusted second extracted parent main population to obtain a thickness adjustment subset. S34. Random individuals are randomly selected from the legal main population to obtain the baseline parent population; Individuals are uniformly and randomly selected from the legitimate main population and the legitimate auxiliary population to obtain the donor population. A cost matrix is ​​constructed between the benchmark parent population and the donor population using the Euclidean distance of the channel thickness vector. Based on a greedy algorithm, the minimum cost pair that has not yet been occupied in the cost matrix is ​​repeatedly selected. The channels of the benchmark parent population are mapped and aligned one by one with the channels of the donor population. A random number of channels in the benchmark parent population are selected to perform whole channel replacement or layer thickness convex combination interpolation and recombination and perturbation is applied to obtain the aligned recombination subset. S35. Merge the main population difference subset, the auxiliary population difference subset, the thickness adjustment subset, and the alignment recombination subset to obtain the candidate offspring population; S36. The candidate offspring population is modified by the manufacturing feasibility constraints and candidate deduplication is performed to obtain the offspring population.

8. A design system for a hyperspectral computational imaging encoder, characterized in that, The design system for the hyperspectral computational imaging encoder includes: The initialization module is used to construct an initial primary candidate population and an initial auxiliary candidate population based on a hybrid initialization strategy, using the thickness of each thin film layer of the filter array as a design variable. The initial primary candidate population and the initial auxiliary candidate population are then modified by manufacturing feasibility constraints to obtain a legal primary population and a legal auxiliary population. The initial primary candidate population and the initial secondary candidate population each include multiple genetic individuals, with each genetic individual corresponding to a group. M×N L The filter thickness matrix, M The number of filters in the filter array. N L The number of thin film layers in a single filter in the filter array, wherein each filter comprises sequentially arranged... N L One gene unit, N L Each of the aforementioned gene units corresponds to a filter layer from the first layer to the second layer. N L The thickness of the thin film layer; The mating fitness calculation module is used to calculate the budget progress based on the current consumed evaluation budget and the preset evaluation budget limit, and to perform overall stagnation detection on the legitimate main population to obtain the stagnation detection result; select the corresponding strategy based on the budget progress and the stagnation detection result, and calculate the mating fitness values ​​corresponding to the legitimate main population and the legitimate auxiliary population; The hierarchical differential module is used to perform hierarchical differential search processing on the legitimate primary population and the legitimate auxiliary population respectively according to the mating fitness value to obtain the offspring population; The optimization target module is used to screen the offspring population based on the ridge regression surrogate model, and to perform optical simulation on each genetic individual in the screened offspring population using the transfer matrix method to obtain the spectral modulation matrix of the corresponding filter array; to establish the optimization target, and to calculate the nominal performance value corresponding to the optimization target based on the spectral modulation matrix; Based on performance constraints, each genetic individual in the offspring population is verified and labeled, and the consumed evaluation budget is updated to obtain a constraint feasibility label. A screening module is used to perform environmental selection on the offspring population based on the nominal performance value and the constraint feasibility marker, thereby selecting the next generation main population and the next generation auxiliary population; wherein, when screening the next generation main population, the selection is based on a preset boundary from the legal main population and the offspring population; when screening the next generation auxiliary population, the selection is based on a dynamic boundary from the legal auxiliary population, the offspring population, and the next generation main population; The judgment module is used to determine whether the updated consumed evaluation budget has reached or exceeded the preset evaluation budget limit: If so, extract all non-dominated solutions from the current next-generation master population to obtain the Pareto optimal filter array design solution set, calculate the peak signal-to-noise ratio corresponding to each candidate design in the Pareto optimal filter array design solution set, and output the candidate design with the largest peak signal-to-noise ratio as the final design scheme of the hyperspectral computational imaging encoder. If not, then the next generation primary population is taken as the new legitimate primary population, and the next generation secondary population is taken as the new legitimate secondary population, and the process returns to the mating fitness calculation module to continue iterative optimization.

9. A computer device, characterized in that, include: The device includes a memory, a processor, and a design program for a hyperspectral computational imaging encoder stored in the memory and executable on the processor. When the processor executes the design program for the hyperspectral computational imaging encoder, it implements the steps in the design method of the hyperspectral computational imaging encoder as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a design program for a hyperspectral computational imaging encoder, which, when executed by a processor, implements the steps in the design method of the hyperspectral computational imaging encoder as described in any one of claims 1-7.

Images