Adaptive anti-glare lighting method

Abstract

This invention discloses an adaptive anti-glare lighting method, aiming to resolve the technical contradiction between measurement efficiency, reconstruction accuracy, and control robustness in dynamic anti-glare control. The method includes: obtaining sparse angular coherence observations through online measurement; rapidly reconstructing the current-cycle angular coherence matrix from the sparse observations based on the inherent low-rank, positive semi-definite structural priors of the angular coherence matrix, and quantifying its angular domain uncertainty; constructing and solving an optimization problem with modal weight increments as decision variables based on the reconstructed matrix and uncertainty, generating coherent modal weights before execution; mapping these weights to an actuator instruction sequence and issuing it, while simultaneously collecting execution records for the next cycle. This invention achieves rapid and accurate sensing and robust closed-loop control of the spatial coherence state of the light field by combining sparse sensing, structured reconstruction, and uncertainty-aware optimization control, thus suppressing dynamic glare.

Description

Technical Field

[0001] This invention belongs to the field of lighting engineering, and in particular to an adaptive anti-glare lighting method. Background Technology

[0002] With the rapid development of semiconductor lighting technology, new light sources, represented by light-emitting diodes (LEDs), have been widely used due to their advantages such as high luminous efficiency, long lifespan, and ease of control. However, while high-brightness light sources enhance the brightness of the lighting environment, they also bring visual health problems, especially dynamic glare. When high-brightness light sources enter the field of vision directly or indirectly, they can cause visual discomfort, decreased vision, or even momentary blindness, creating a conflict between lighting function and visual safety in specific work scenarios (such as night driving, precision surgery, and security monitoring). Therefore, research into adaptive lighting technology that can eliminate glare at the center of the field of vision in real time and with precision based on the dynamic changes in the observer's line of sight can optimize the adaptability of lighting systems to the light environment, solve visual safety issues in the application of high-brightness light sources, and improve the technical adaptability of lighting systems in specific work scenarios.

[0003] Current research on glare issues primarily focuses on suppressing glare through optical design or simple zone control. In terms of optical design, passive optical elements such as frosted or microstructured diffusers, lens arrays, and grids are typically used to expand the light-emitting area, homogenize the light emission angle, and reduce local brightness on the light source surface. Regarding active control, some lighting systems divide the light-emitting panel into multiple independently controllable zones. Cameras or sensors track the user's eye position, and when the system detects the gaze entering a particular zone, it lowers the brightness of that zone or shuts it off entirely. Furthermore, some research is exploring the use of wavefront shaping techniques, employing devices such as spatial light modulators to preliminarily control the light field, creating dark areas in specific spaces to avoid direct illumination.

[0004] However, when dealing with complex and dynamic anti-glare applications, achieving high-fidelity, real-time sensing and predictive control of the physical state of the light field remains a challenge. Current technologies rely on indirect and coarse-grained control of light intensity, lacking means to perceive the angular coherence matrix—a core physical property of the light field—in real time, thus limiting the precision and degrees of freedom of anti-glare control. Therefore, further research and innovation are needed to address these issues in existing technologies. Summary of the Invention

[0005] Purpose of the invention: In view of the above-mentioned problems of the prior art, this application provides an adaptive anti-glare lighting method.

[0006] Technical solution: According to one aspect of this application, an adaptive anti-glare lighting method includes:

[0007] Based on the line-of-sight cone, the engineering index boundary, and the previous period's angular mutual coherence matrix and the previous period's coherent mode weight vector, sparse angular mutual coherence observations are obtained through online measurement.

[0008] Based on the sparse angular mutual coherence observations and the angular mutual coherence matrix of the previous period, the angular mutual coherence matrix of the current period and the corresponding angular domain uncertainty are reconstructed.

[0009] Based on the current period's angular coherence matrix, angular domain uncertainty, and the previous period's coherent mode weight vector, optimization is performed using the mode weight increment as the decision variable to obtain the coherent mode weights before execution.

[0010] The coherent mode weights before execution are mapped to the executor instruction sequence and issued. At the same time, execution records are collected to update the previous cycle angular coherence matrix and the previous cycle coherent mode weight vector for the next cycle.

[0011] Beneficial effects: This invention introduces online reconstruction of angularly coherent matrices and angular domain uncertainty quantification to achieve high-fidelity perception of the physical state of the light field; combined with modal weight incremental optimization and periodic iterative updates, predictive control is achieved. It solves the bottleneck of dynamic scene adaptation in existing technologies, improves the real-time performance and accuracy of anti-glare response, and ensures lighting safety. The related technical effects will be described in detail below with reference to specific embodiments. Attached Figure Description

[0012] Figure 1 A flowchart of the adaptive anti-glare lighting method provided in the embodiments of this application.

[0013] Figure 2 This is a flowchart illustrating the online measurement process for obtaining sparse angular mutual coherence observations and reconstructing the angular mutual coherence matrix and angular domain uncertainty for the current period, as provided in the embodiments of this application.

[0014] Figure 3 A flowchart for planning the sampling layout and baseline selection is provided for the embodiments of this application.

[0015] Figure 4 This is a flowchart illustrating the optimized generation of coherent mode weights before execution, provided in an embodiment of this application.

[0016] Figure 5 This is a flowchart of another optimized generation of coherent mode weights before execution, provided as an embodiment of this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0018] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] To address the aforementioned issues, the applicant conducted in-depth searches and analyses, and discovered:

[0020] Glare is fundamentally related to the spatial coherence of the light field. However, current technologies mostly rely on indirect and coarse-grained control of light intensity, lacking a means to perceive the angular coherence matrix—a core physical property of the light field—in real time, thus limiting the precision and freedom of anti-glare control. Furthermore, even with direct measurement, the complete angular coherence matrix is ​​a high-dimensional object, and the time required for full acquisition far exceeds the human eye's response time, leading to a conflict between measurement completeness and system real-time performance. While sparse measurement can improve efficiency, accurately recovering the complete matrix from undersampled data remains an ill-posed inverse problem.

[0021] Based on this, the reconstruction results based on sparse and noisy observations are accompanied by uncertainties. If the controller ignores the model uncertainty, its output instructions may not be able to stably achieve the expected performance indicators in the real physical world, resulting in inconsistent anti-glare effects and a lack of robustness.

[0022] To solve these problems, combined with Figures 1 to 5 The present invention will be specifically described through the following embodiments.

[0023] It should be understood that, in order to make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1: An adaptive anti-glare lighting method is provided, which can be implemented in a lighting device including a light source system, a control system, and at least a gaze tracking sensor.

[0025] Specifically, the light source system can be a complex light source capable of dynamically controlling the spatial coherence of the emitted light field, such as a light source based on a spatial light modulator (SLM), a digital micromirror device (DMD), or a multi-core fiber array. The control system, such as an embedded computing unit or an industrial computer, is responsible for executing the method of this invention. A gaze-tracking sensor is used to acquire the observer's gaze direction in real-time or near real-time.

[0026] In the context of this invention, to ensure consistency in description, technical terms and data items are defined throughout the various embodiments, as follows:

[0027] Line-of-sight cone (Cone): A narrow-angle three-dimensional region representing the current observer's gaze direction, typically with a half-angle between 2 and 5 degrees; within this cone, spatial coherence of the light field is suppressed to eliminate glare. Angular coherence matrix J: A Hermitian matrix describing the coherence relationship between light sources at different exit angles; its diagonal elements represent the light intensity at each angle, and its off-diagonal elements J... _ij The cross-coherence between angles i and j represents the degree of cross-coherence; it serves as the basis for subsequent analysis, reconstruction, and optimization. For example, in a system discretized into N angles, the angular cross-coherence matrix is ​​an N×N matrix. The coherent mode weight vector w: The overall coherent state of the light source can be decomposed into a linear superposition of a series of basis coherent modes; each component of this vector is the weight coefficient of each basis mode. The control system adjusts this vector to control the coherent state of the light field. Previous cycle / Current cycle: The method of this invention operates in a periodic iterative manner, for example, every 50 milliseconds or 100 milliseconds; the previous cycle refers to the state or data of the previous control cycle, while the current cycle refers to the control cycle currently being calculated and executed.

[0028] Sampling Layout and Baseline Selection Plan: The channel and corner pair combination scheme for sparse sampling in this period. Sparse Angular Inter-coherence Observation y: A small number of corner pair inter-coherence sample values ​​collected based on the sampling layout. Angular Inter-coherence Matrix J from the previous period. _prev : Estimation of the angular mutual coherence principal structure of the previous period. The angular mutual coherence matrix J of this period.# The angularly interdependent principal structure estimated by rapid reconstruction in this cycle, where... # Represents the hat symbol. Angular domain uncertainty U # : To J # Uncertainty assessment at each angular location. Angular complex coherence amplitude profile Γ # : By J # Derived angular complex coherence amplitude distribution. Groove depth estimation D. _notch # : Coherence groove depth estimation for the view cone Cone; further, D _notch The groove depth. Uncertainty-weighted profile Gw # : Angular domain uncertainty U # propagation to Γ # The subsequent weighted profile. Engineering index boundary Spec. _limit Constraints include: illuminance, color rendering index or TM-30 index, flicker SVM, energy efficiency, etc. Target constraint set. _set : The target set containing the groove depth threshold and the allowable response time. Structural priors: Contains structured prior parameters such as low rank, near Toeplitz, band-limited, and positive semidefinite.

[0029] Upper periodic coherent mode weight vector w _prev : Coherent mode weights from the previous cycle. Coherent mode weights w before execution. _exec : Coherent mode weights used in actual execution after comprehensive constraints. Current period coherent mode weights w _new : Candidate coherent mode weights for the current period obtained through optimization. Mode weight increment Δw: w _new With w _prev The difference components satisfy w _new =w _prev +Δw. Constraint Map _cons Approximate mapping from coherent mode weights to indices such as illuminance, color rendering index, and flicker. Feasible region. _set The feasible region is determined by the actuator speed and amplitude boundaries and engineering constraints. Calib_map: Factory calibration mapping from coherent mode weights to actuator instruction sequences. Actuator instruction sequence u: A set of instructions including pump current allocation, SLM phase diagram, and DMD angle multiplexing timing. Execution record R _exec Telemetry and summarization are performed in this cycle and used for initialization in the next cycle.

[0030] In this embodiment, the adaptive anti-glare lighting method includes line-of-sight perception and constraint convergence, which may specifically include the following steps:

[0031] Step S1.1: The control system reads raw data from the gaze tracking sensor to determine the gaze cone (Cone) for the current cycle; and combines this with the system-preset or dynamically input engineering parameter boundary (Spec) from external environmental sensors (such as a luminance meter). _limit (For example, requirements include a total illuminance of no less than 500 lux, a color rendering index (Ra) of no less than 90, and a flicker metric (SVM) of less than 0.4, etc.), and the current state of the system (such as response speed requirements), to generate the target constraint set Target. _set This set defines the performance targets to be achieved in this cycle, which may include, for example, a minimum threshold for the groove depth (i.e., the degree to which coherence within the view cone needs to be suppressed) and a maximum allowable response time window (e.g., adjustment must be completed within 150 milliseconds). Accordingly, the system loads the previous cycle's angular coherence matrix J from the previous cycle's execution record. _prev The weight vector w of the coherent mode of the previous period _prev As the initial state.

[0032] Step S1.2: The control system reads the previous period's angular mutual coherence matrix J. _prev Based on the known physical model, a structural prior, Prior_struct, is constructed. This structural prior is a mathematical description of the structured characteristics that the true angularly coherent matrix should satisfy, used to guide sparse reconstruction and sampling planning in subsequent steps. In a preferred embodiment, the structural prior includes at least three properties selected from the group consisting of: low-rank property of the angularly coherent matrix, near-Toeplitz property, band-limited property, and positive semi-definite property.

[0033] Among these characteristics, the positive semi-definite property (PSD) is crucial. As the covariance matrix describing the energy distribution of a physical system, the angular mutual coherence matrix J must be positive semi-definite, meaning all its eigenvalues ​​are non-negative. The low-rank property is important because in lighting scenarios, the spatial coherence of the light field is mainly determined by a few dominant coherent modes, making the angular mutual coherence matrix J typically low-rank or well approximated by a low-rank matrix. The near-Toeplitz property is also important because for spatially uniform or quasi-uniform light sources, the coherence between different angles depends primarily on the angle difference rather than the absolute angular position, resulting in a near-Toeplitz structure for the angular mutual coherence matrix J, where the element values ​​along the diagonal direction are approximately equal. Finally, the band-limited property is crucial because the physical aperture of the light source is finite, making its angular domain spectrum (obtained through a Fourier transform of the diagonal mutual coherence matrix) band-limited, meaning high-frequency components are zero or negligible. Based on these characteristics, the system completes all preparatory work before online measurement and optimization, generating the target constraint set Target. _set The prior structure, Prior_struct, provides the target and mathematical tools for subsequent steps.

[0034] Example 2 provides a specific implementation scheme for the adaptive anti-glare lighting method, used to provide a general description of the technical solution of the present invention; it includes the following steps:

[0035] In one example, sparse angularly coherent observations are obtained online based on the line-of-sight cone, engineering index boundaries, and the previous-period angularly coherent matrix and coherent mode weight vector from the previous period. In some embodiments, the system integrates the information from the line-of-sight cone, engineering index boundaries, and the previous-period angularly coherent matrix and coherent mode weight vector from the previous period to formulate a sampling scheme and measure sparse angularly coherent observations online.

[0036] Specifically, at the start of the control cycle of this invention, the system gathers the line-of-sight cone (Cone) from the line-of-sight sensor and the system-required engineering specification boundary (Spec). _limit And the upper-period angular coherence matrix J, which serves as historical information. _prev The weight vector w of the coherent mode of the previous period _prev Based on this, a sampling scheme is planned, which specifies which angle pairs' mutual coherence needs to be measured within the current period. Since measuring the mutual coherence of all angle pairs is too time-consuming and cannot meet real-time requirements, this sampling scheme is sparse, meaning it only selects a small subset of all possible angle pairs for measurement. Accordingly, the control system drives the light source system (e.g., by controlling a specific channel combination of multi-core optical fibers or a specific area on the SLM) to perform measurements according to this sparse sampling scheme, obtaining a limited number of sample values ​​but with a high information content, i.e., sparse angular mutual coherence observations y.

[0037] Furthermore, based on sparse angular coherent observations and the angular coherence matrix of the previous period, the angular coherence matrix of the current period and its corresponding angular domain uncertainty are reconstructed. According to one aspect of this application, after obtaining sparse angular coherent observations through online measurement, the angular coherence matrix of the current period and its corresponding angular domain uncertainty are optimized by combining the angular coherence matrix of the previous period.

[0038] Accordingly, after obtaining the sparse angularly coherent observation y, the system needs to recover the complete current-period angularly coherent matrix J describing the current optical field state from the fragmented information. # This belongs to the compressed sensing or matrix filling problem. This step utilizes the structural prior `Prior_struct` (containing properties such as low rank, near-Toeplitz, band-limited, and positive semi-definite) as constraints, based on the J-value of the previous period. _prev As the starting point or regularization term of the iteration, the matrix that satisfies all structural priors and matches the actual observed value y is found through optimization of the solution process, and is taken as the angular coherence matrix J of the current period. #The estimated value. During the reconstruction process, the system also simultaneously evaluated the reliability or confidence level of the reconstruction result at each angular location, forming a quantified angular uncertainty U. # U # In some angular regions, the values ​​are high, and due to the lack of direct observation or weak structural prior constraints, the reconstruction results in these regions have greater uncertainty.

[0039] Furthermore, based on the current period's angular coherence matrix, angular domain uncertainty, and the previous period's coherent mode weight vector, optimization is performed using the mode weight increment as the decision variable to obtain the coherent mode weights before execution. The current optical field state J is then reconstructed. # Then, the method for adjusting the light source to achieve the anti-glare effect is calculated. Accordingly, an optimization model is established to minimize the residual coherence within the viewpoint cone (Cone) while satisfying all engineering constraints (such as brightness and color rendering). The decision variable in this model is not a completely new coherent mode weight vector w, but rather the mode weight increment Δw, i.e., the change in weight in the current cycle relative to the weight in the previous cycle. This allows for a smaller optimization problem size, faster solution, and direct constraints on the rate and magnitude of weight changes, ensuring a smooth transition. In the objective function and constraints of the optimization model, the system comprehensively considers J... # The derived coherence profile, U # The indicated uncertainty (applying a more conservative strategy to areas of high uncertainty), and w _prev The initial state. After solving this optimization problem, the optimal modal weight increment Δw is obtained, and compared with w. _prev Superposition (w) _new =w _prev +Δw), after safety verification and smoothing, obtain the coherent modal weights w before execution. _exec .

[0040] Based on this, the pre-execution coherent mode weights are mapped to an actuator instruction sequence and issued. Simultaneously, execution records are collected to update the previous cycle angular coherence matrix and the previous cycle coherent mode weight vector for the next cycle. The system calculates the abstract pre-execution coherent mode weights w. _exec The system uses a pre-calibrated mapping model (Calib_map) to convert data into specific instructions that the light source hardware can directly execute, i.e., an actuator instruction sequence u. This set of instructions may include the current distribution scheme for the pump laser, the phase pattern to be loaded for the SLM, or the timing control signals for the DMD, etc. These instructions are then sent to the light source driver hardware for execution, completing the adjustment of the coherent state of the optical field. Simultaneously with or after instruction execution, the system collects necessary telemetry data (such as actual output light intensity, drive current, etc.) to form an execution record R. _exec This record, along with the J calculated for this period, # and w_exec This will be stored as the upper-cycle angular coherence matrix J at the start of the next control cycle. _prev The weight vector w of the coherent mode of the previous period _prev This forms a closed-loop, continuously adaptive control process. Therefore, this invention can respond in real time to changes in the observer's line of sight, dynamically creating a low-coherence groove in the center of the line of sight to suppress glare, while maintaining the lighting quality and overall energy efficiency of the surrounding area, thus achieving adaptive healthy lighting.

[0041] Alternatively, the adaptive anti-glare lighting method can also be implemented using the following steps:

[0042] Step S2.1, line-of-sight perception and constraint convergence; read the line-of-sight cone (Cone) and engineering specification boundary (Spec). _limit Upper periodic angular mutual coherence matrix J _prev Upper period coherent mode weight vector w _prev By combining the system clock and sensor status, the minimum threshold for groove depth and the allowable response time window are calculated to obtain the target constraint set Target. _set Read the angular coherence matrix J of the previous period. _prev Upper period coherent mode weight vector w _prev We introduce structural assumptions such as low rank, near Toeplitz, band-limited and semi-positive definite to form the structural prior_struct, which is used to guide subsequent sampling planning and fast reconstruction.

[0043] Step S2.2: Sparse Angular Intercoherence Online Measurement and Fast Reconstruction; Read the line-of-sight cone (Cone) and the structural prior (Prior_struct). Based on the sensitivity of the groove edge to the objective function and the historical error distribution, generate a sampling scheme that prioritizes coverage of the groove edge and the high-uncertainty angular domain, obtaining the sampling layout and baseline selection plan. Control the small aperture or multi-core fiber channel according to the sampling layout and baseline selection plan, collect a small number of angular intercoherence samples, and perform normalization, timestamp, and noise estimation to obtain the sparse angular intercoherence observation y (including quality labels). Read the sparse angular intercoherence observation y, the structural prior (Prior_struct), and the upper-period angular intercoherence matrix J. _prev Under the joint priors of low rank, near Toeplitz, band-limited and positive semidefinite, one or a finite number of projection iterations are performed to obtain the periodic angular coherence matrix J. # With angular domain uncertainty U # From the current period, the mutual coherence matrix J # Calculate the angular complex coherence amplitude profile Γ # And calculate the groove depth estimate D for the line-of-sight cone Cone. _notch # ; angular domain uncertainty U# Propagation to the angular complex coherence amplitude profile and groove depth estimation forms the uncertainty weighted profile Gw. # .

[0044] Step S2.3: Coherence mode weight prediction optimization and constraint fusion; read the angular coherence matrix J for this period. # Uncertainty-weighted profile Gw # Target constraint set _set Engineering indicator boundaries Spec _limit Upper period coherent mode weight vector w _prev The approximate mapping from coherent mode weights to illuminance, color rendering, and flicker is updated in small steps to obtain the constraint mapping Map. _cons Based on constraint mapping, actuator rate and amplitude limits, the feasible region is determined, resulting in the feasible region (Feasible). _set Read the angular complex coherence amplitude profile Γ # Uncertainty-weighted profile Gw # Target constraint set _set Feasible domain _set Upper period coherent mode weight vector w _prev Using the modal weight increment Δw as the decision variable, a small-scale short-time-domain optimization model is constructed to minimize the in-cone coherence cost and engineering index deviation, while penalizing excessively rapid weight changes and incorporating uncertainty weights. This can be described by the following formula:

[0045] min Δw L _notch (Δw;Γ # Cone)+λ _1 L _spec (Δw;Map _cons Spec _limit )+λ _2 L _Dyn (Δw;Target _set )+λ _3 L _uncert (Δw;Gw # );stD _notch (Δw;Γ # Cone)≥D _thresh , Δw∈Feasible _set w _new =w _prev +Δw;

[0046] Where, λ _1 , λ _2、 λ _3The weighting coefficients for each cost are used to balance the importance of different performance indicators in the cost function L. _notch The cost function L measures the cumulative magnitude of angular complex coherence within the line-of-sight cone (Cone). _spec Through constraint mapping Map _cons Approximate illuminance, color rendering index or TM-30 index, flicker SVM and energy efficiency deviation; cost function L _Dyn Based on the target constraint set Target _set Give the response time-related weights; cost function L _uncert Using uncertainty weighted profile Gw # Increased cost for high-uncertainty angular domains; Groove depth threshold D _thresh From the target constraint set Target _set Solving the optimization problem yields the coherent mode weights w for the current period. _new Combined with the previous periodic coherent mode weight vector w _prev To meet the requirements of smooth switching of actuators, for w _new Perform safety incremental correction to obtain the coherent mode weights w before execution. _exec .

[0047] Step S2.4: Coherent state execution and energy equivalence mapping; read the coherent mode weights w before execution. _exec The Calib_map is calibrated to generate energy-equivalent phase and mode switching sequences under the principle of minimum intensity and spectral perturbation, resulting in the actuator command sequence u (including pump current allocation, SLM phase map, and DMD angle multiplexing timing). The actuator command sequence u is then sent to the hardware to complete the coherent state switching for the current cycle; necessary low-overhead operational telemetry is collected to form the execution record R. _exec In order to transform the angularly interdependent matrix J of the next period into the previous period. _prev The weight vector w of the coherent mode of the previous period _prev The initialization basis.

[0048] Example 3 provides an optional implementation of lighting control based on structural priors and incremental optimization, used to optimize the generation of coherent modal weights before execution. This example is based on an optimization framework with modal weight increments as decision variables, incorporating groove cost, engineering index cost, and dynamic response cost. The application of this technical solution requires consideration of the inherent structural prior characteristics of the diagonal coherent matrix.

[0049] In a specific implementation, the optimization of generating coherent mode weights before execution is achieved by constructing and solving an optimization problem with mode weight increments as decision variables; in other words, the coherent mode weights before execution are generated by using and solving an optimization problem with mode weight increments as decision variables. The coherent mode weights for the current period are formed by superimposing the coherent mode weight vector from the previous period with the solved mode weight increments. Furthermore, the decision variables of this optimization problem are not the final coherent mode weight vector w for the current period. _new It is not itself, but rather its relative to the previous period's weight vector w. _prev The increment Δw, where w satisfies the relation _new =w _prev Using incremental Δw as the decision variable has at least the following technical advantages: Firstly, it anchors the search space of the optimization problem near the solution of the previous cycle, enabling rapid convergence in scenarios with high real-time requirements. Secondly, it allows direct constraints on the norm or components of Δw, limiting the rate and magnitude of change of the weight vector, ensuring smooth transitions between coherent states, avoiding excessive impact on the actuator, and preventing flickering that may be perceptible to the human eye. Thirdly, it makes the modeling of dynamic response costs more intuitive and accurate.

[0050] Furthermore, the optimization problem includes a cost term related to response time. The determination of this cost term involves mapping the modal weight increment to the expected execution time based on the actuator rate and amplitude boundaries defined in the feasible region, and penalizing the portion of the expected execution time that exceeds the preset response time upper limit in the target constraint set. In the optimization model, the system converts any candidate Δw into the expected physical execution time T according to the preset actuator model. _pred For example, the model could be a piecewise linear or lookup table relationship established based on the refresh rate and maximum phase modulation depth of the actuator (such as an SLM), T _pred =f(Δw). If the calculated T _pred Exceeding the target constraint set Target _set The upper limit of response time T set in the middle _resp_max Then the penalty term, for example, γ*max(0, T) _pred -T _resp_max This will be added to the overall objective function, where * is the multiplication operator and γ is the penalty weight coefficient. Based on this, the optimizer will automatically find a compromise solution that satisfies the anti-glare requirements while completing the adjustment within a specified time.

[0051] In a preferred embodiment, the objective function to be minimized during the optimization process is composed of at least three weighted costs: a groove cost calculated based on the current period's angular coherence matrix to penalize residual coherence within the line-of-sight cone; an engineering index cost calculated based on the engineering index boundary to penalize deviations of photometric or electrical indices from their preset range; and a dynamic response cost calculated based on the objective constraint set to penalize the rate or magnitude of change in coherent mode weights. This objective function can be expressed as a weighted sum in the following form: min _Δw L _notch +λ _1 *L _spec +λ _2 *L _Dyn ;

[0052] Where, λ _1 and λ _2 These are the weighting coefficients for each cost, used to balance the importance of different performance indicators. Groove cost L _notch This term is used to form the deepest possible coherence groove within the line-of-sight cone Cone; it is based on the angular coherence matrix J of the current period. # The derived angular complex coherence amplitude profile Γ # The calculations are performed. To predict the impact of Δw on coherence during optimization, the system employs a fast approximation update rule, such as Γ. _pred (Δw)≈Γ # +J _Γ *Δw, where Γ _pred It predicts coherence; J _Γ It is the Jacobian matrix, representing the local sensitivity of coherence to mode weights. L _notch It is then constructed as a weighted accumulation of the predicted coherence magnitude at discrete angle points θ within the view cone Cone, i.e., L _notch =∑(ρ(θ)*|Γ _pred (θ)|), where ρ(θ) is the angle weight, which can be set to have a larger value at the edge of the groove to sharpen the groove boundary.

[0053] Engineering index cost L _spec This ensures that lighting quality is not degraded by anti-glare adjustments, utilizing a constraint mapping (Map). _cons This is used to approximately predict the impact of Δw on various engineering indicators (such as total illuminance, color rendering index, flicker SVM, etc.). If a certain predictable indicator... _k Exceeding the engineering specification boundary. _limit The specified range, penalty function (e.g., quadratic penalty or hinge loss function) penalty(metric) _k -spec _k ) will be activated and counted in L _spec; where metric _k As a predictive indicator, spec _k This represents the boundary of the range.

[0054] Dynamic response cost L _Dyn As mentioned before, this item is directly related to Δw. In addition to the aforementioned penalty related to response time, L... _Dyn It can also include the norm of Δw (e.g., ||Δw||2). 2 The penalty term () is a form of regularization and can also guide the optimizer to choose a solution with a more gradual change.

[0055] It should be noted that the above optimization process is implemented to obtain the angular coherence matrix J for this period. # This is a prerequisite. Further reconstruction requires incorporating relevant structural prior properties. In this invention, structural priors include at least three properties selected from the group consisting of: low-rank property of angularly coherent matrices, near-Toeplitz property, band-limited property, and positive semi-definite property. In other words, structural priors include at least three of the following properties: low-rank property of angularly coherent matrices, near-Toeplitz property, band-limited property, and positive semi-definite property. It should be understood that prior knowledge is used as a constraint in the reconstruction step to recover the complete matrix from a few observations. For example, in the reconstruction algorithm, the positive semi-definite and low-rank properties can be enforced by performing eigenvalue decomposition on the intermediate iteration result matrix and discarding small non-negative eigenvalues; the band-limited property can be enforced by performing low-pass filtering on the matrix in the Fourier domain; and the near-Toeplitz property can be enforced by averaging along the diagonal direction.

[0056] In some alternative implementations, the cost term in the above optimization model can take different mathematical forms. For example, L _notch It can be defined as the maximum coherence within the line-of-sight cone (Cone), used to modulate the suppression effect under worst-case conditions. _spec The penalty in the equation can also be implemented as a hard constraint, which prohibits any solution that exceeds the boundary, but may lead to an unsolvable optimization problem under certain critical conditions. Furthermore, the entire optimization problem can be modeled as linear programming (LP), quadratic programming (QP), or the more complex second-order cone programming (SOCP) depending on the specific form of the cost and constraints, and can be solved efficiently using the corresponding standard solvers.

[0057] An alternative implementation is to use incremental short-time domain optimization modeling as the basic model for optimization. This model clarifies the specific mathematical construction of each cost term; the objective function for solving the optimal modal weight increment Δw is composed of a weighted sum of four costs:

[0058] min _Δw L _notch(Δw)+λ _1 *L _spec (Δw)+λ _2 *L _Dyn (Δw)+λ _3 *L _uncert (Δw); hard constraints also need to be satisfied: D _notch (Δw)≥D _thresh And Δw∈Feasible _set ;

[0059] Among them, the groove cost term L _notch Discretize the line-of-sight cone Cone into angle Θ _cone ; through linear approximation Γ _pred (Δw)≈Γ # +J _Γ *Δw predicts changes in coherence. The cost function is defined as a weighted sum of the predicted coherence magnitudes within the cone: L _notch (Δw)=∑ _θ∈Θ_cone ρ _cone (θ)*|Γ _pred (θ)|, where ρ _cone (θ) is the angle weight. This is achieved using a constraint mapping (Map). _cons Predict the changes in various engineering indicators caused by Δw; the cost function is the value of each indicator exceeding its boundary Spec. _limit The sum of penalties: L _spec (Δw)=∑ _k β _k *penalty(metric _k (Δw)-spec _k Map Δw to the expected execution time T using the actuator model. _pred (Δw); The cost function is the response time limit T. _resp_max Partial punishment: L _Dyn (Δw)=γ*max(0,T) _pred (Δw)-T _resp_max Using the uncertainty-weighted profile Gw # As a weight, a weighted penalty is applied to the predicted coherence: L _uncert (Δw)=∑ _θ ρ _uncert (θ)*|Γ _pred (θ)|, where ρ _uncert (θ) by Gw # The given value represents its weight. This term causes the optimizer to be more inclined to suppress coherence in regions of high uncertainty. This optimization problem can be transformed into a quadratic programming (QP) or second-order cone programming (SOCP) problem, which can be solved in milliseconds using existing mature numerical optimization solvers.

[0060] Example 4 describes a preferred implementation of the online reconstruction method based on sequential coupled projection. In this example, the reconstruction process is achieved by adaptively adjusting the prior structural parameters in a data-driven manner.

[0061] According to this embodiment, obtaining sparse angular intercoherent observations through online measurement and reconstructing the current period's angular intercoherence matrix and angular domain uncertainty can be achieved as follows: A sampling layout and baseline selection are generated based on maximizing the identifiability of the structural parameters inherent in the prior structure; sparse angular intercoherent observations are collected based on the sampling layout and baseline selection; the current period's angular intercoherence matrix is ​​reconstructed using a sequential coupled projection method, wherein the parameters resolved in one structural prior projection are used to adaptively adjust the constraints of another subsequent structural prior projection. In other words, the sampling layout and baseline selection are planned based on identifiability; the identifiability is correlated with maximizing the structural parameters inherent in the prior structure; sparse angular intercoherent observations are collected accordingly; next, based on the structural prior, the current period's angular intercoherence matrix is ​​reconstructed using a sequential coupled projection method, wherein the parameters resolved in the projection are used to adaptively adjust the constraints of another subsequent structural prior projection.

[0062] For example, a structure-adaptive sampling plan is performed. This stage is used to generate a sampling layout and baseline selection plan that maximizes information gain. Unlike dense sampling in regions of interest (such as line-of-sight pyramids), this method uses maximizing the identifiability of structural parameters contained in the structural prior as the criterion. These structural parameters refer to the specific values ​​describing the structural prior, Prior_struct, such as the rank K in a low-rank prior, the bandwidth in a band-limited prior, and the stationary scale in a near-Toeplitz prior.

[0063] In specific implementation, the system is based on the upper periodic angular mutual coherence matrix J. _prev A discriminability field or parameter sensitivity map, Id_map, is constructed. This map assigns a score to each candidate corner pair (i, j) to be sampled, which measures the contribution of the observation for that corner pair to the accurate identification of the aforementioned structural parameters. For example, an approximate Fisher information can be used as a metric; the larger the Fisher information, the smaller the lower bound of the variance of the observation for parameter estimation, i.e., the higher the discriminability. Further, the system sorts all candidate corner pairs according to the score / cost ratio, and, combined with the observation matrix condition number constraint (avoiding the selection of linearly correlated observations) and the corner domain coverage balance constraint, selects a subset of corner pairs with the highest scores to form the sampling layout and baseline selection Plan. Based on this, the sampling process shifts from passive observation to active exploration, used to collect highly correlated information related to the model, providing data support for subsequent reconstruction operations.

[0064] Another example, structural adaptive sampling planning (aiming at identifiability) can also be implemented using the following method:

[0065] Read the line-of-sight cone Cone and the structural prior Prior_struct (including low-rank range, near-Toeplitz stationary scale, band-limited frequency band, and positive semi-definite projection radius), and generate a structural parameter sensitivity field on the angular domain grid. Specifically, the above periodic angular coherence matrix J... _prev For reference, considering parameters such as low-rank order K, band-limited boundaries, and near-Toeplitz windows, the first-order sensitivity of each corner pair from observation error to parameter error is calculated, forming a saliency distribution Id_map (corner pair to scalar score) for channel selection scoring. The saliency distribution Id_map and the channel geometry and aperture constraints in the structural prior Prior_struct are read. Combining the line-of-sight cone Cone, priority is given to cone edges and high-uncertainty corner regions. Channels are selected using the following rules: sorted by score / cost, the score of each candidate corner pair is the saliency score divided by the sampling cost (timing budget, switching loss, geometric reachability). Geometric coverage and condition number constraints are applied, requiring the selected corner pairs to have a good geometric distribution in the corner region, so that the condition number of the observation matrix is ​​controlled during reconstruction. A candidate channel list Cand_list (with sorting and constraint flags) is obtained. The candidate channel list Cand_list and the system clock budget are read, and actual sampling quotas and trigger timings are allocated to match the switching cycle of the execution end; a sampling layout and baseline selection Plan are generated.

[0066] The sampling layout and baseline selection plan are read, and offset correction and amplitude / phase calibration are performed on the selected channels to obtain the channel calibration parameter Calib_chan. Acquisition is triggered according to the sampling layout and baseline selection plan, and amplitude and phase are normalized using the channel calibration parameter Calib_chan to obtain the sparse angular inter-coherence observation y. _raw (Including timestamps and channel identifiers). For sparse angularly mutually coherent observations y _raw Estimate noise power and packet loss flags, and output sparse angular interference observation y (with quality label).

[0067] Read the sparse angularly coherent observations y and the sampling layout and baseline selection plan, and construct the observation operator: for each selected angle pair (θ) _i θ _J Generate the index embedding matrix E. _ij The observation operator P is obtained by weighting the aperture function and geometric baseline length according to the prior structure Prior_struct; and by superimposing the real and imaginary parts of the phase geometry terms. _Ω (Sparse linear mapping). Read the upper-period angular coherence matrix J. _prev Scale registration is performed between the angle grid and the strength reference to obtain the initial value J._0 ; For J _0 Performing positive semidefinite projection (spectral clipping and trace normalization) yields the positive semidefinite matrix J. _psd For a semi-positive definite matrix J _psd The low-rank projection is performed, and the rank threshold is jointly determined by the discriminability noise estimation, outputting a low-rank matrix J. _lr With rank estimation K _est Read the low-rank matrix J _lr With rank estimation K _est Perform near-Toeplitz smoothing (window by K) _est (Adaptive), then perform band-limited frequency band clipping (the frequency band is determined by the prior structure Prior_struct and J). _lr (The spectral energy distribution is jointly determined), thus obtaining the intermediate term J. _bl ; Apply the observation operator P _Ω J _bl Least-square back projection is performed between the observed location and the sparse angular coherent observation y to generate the observation consistency matrix J. _cons For the observation consistency matrix J _cons Two types of convergence are checked: the residual norm and the difference between adjacent projections are observed, and the process terminates if they meet the criteria; based on the residual spectral distribution, structural deviation, and angular domain energy leakage rate, the angular domain uncertainty U is generated. # J _cons Let J be the angular coherence matrix of this period. # .

[0068] Furthermore, based on the plan generated in the previous stage, the sparse observation acquisition and labeling system drives the hardware to perform online measurements, acquiring sparse angularly interdependent observations y. This process can further include channel self-calibration and zero-point correction to improve the accuracy of the observation data. The acquired raw data y... _raw It will undergo purification and quality labeling processes, such as estimating the signal-to-noise ratio of each observation or adding quality labels, to generate sparse angularly coherent observations y with quality information for use in the reconstruction stage.

[0069] Furthermore, a sequential coupled projection method is used to reconstruct the angular coherence matrix of the current period. Unlike methods that treat all structural priors (low-rank, positive semi-definite, etc.) as fixed constraints and apply them simultaneously, this method organizes them into an ordered, interdependent projection chain. Furthermore, the parameters resolved in the structural prior projection are used to adaptively adjust the constraints of subsequent structural prior projection steps. Accordingly, the sequential coupled projection process is as follows:

[0070] With J _prev Let J be the initial value. _0 Performing a positive semidefinite projection, i.e., cropping negative eigenvalues ​​to zero or small positive numbers τ through eigenvalue decomposition. _psdThe intermediate matrix J is obtained. _psd For J _psd The algorithm performs low-rank projection; it dynamically estimates the most likely rank K based on the energy distribution of the singular values. _est Among them, K _est This is not only used in this low-rank projection, but will also be passed as an information parameter to subsequent steps. Next, the matrix J after low-rank projection... _lr Perform near-Toeplitz smoothing and band-limited projection; where the window size used for near-Toeplitz smoothing is based on the rank K estimated in the previous step. _est Adaptively adjusted; and the band boundary of the band-limited projection can also be determined according to J. _lr The spectral energy distribution and known stationary scale are inversely extrapolated; it should be understood that this information transfer breaks the independence between priors, making them work together. After applying a round of structural prior constraints, the resulting matrix may no longer precisely satisfy the collected observation value y; therefore, a backpropagation step is required, that is, at the position of the observed entry, the value of the matrix is ​​forcibly corrected to the observation value y, or the solution closest to y is found through the least squares method to obtain J. _cons The above sequential coupled projection process can be iterated a finite number of times until the observed residual or the change in the matrix between two adjacent iterations is less than a preset threshold, at which point the angular coherence matrix J for the current period is output. # .

[0071] According to this embodiment, in the implementation of sequential coupled projection, the order of projection is not fixed and can be adjusted according to the prior strength of the specific application scenario. As an optional implementation, for example, if the band-limiting characteristic is very strong, it can be used as the first projection step. Furthermore, the coupling method of the parameters can also be diversified, for example, rank estimation K... _est This can directly affect the threshold of band-limited projection. The design of this parameter being passed in the projection chain and adaptively adjusting subsequent constraints is a distinguishing feature from conventional projection algorithms in terms of technical solution.

[0072] Another example is that the reconstruction step can also be implemented using the following process: the matrix to be reconstructed is sequentially projected onto a set of constraints defined by multiple structural priors, and the reconstruction result is quickly obtained within a finite number of iterations (usually one or a few steps). The complete reconstruction cycle specifically includes: based on the sampling layout and baseline selection Plan, the system constructs a linear observation operator P. _Ω This operator is used to extract the matrix elements corresponding to sparse angularly coherent observations y from the complete angularly coherent matrix J. The previous period's angularly coherent matrix J... _prev J serves as the initial value for this reconstruction iteration. _0 To ensure consistent scaling, J can be scaled first. _prev Perform normalization or scale registration.

[0073] For the current iteration matrix (initially J) _0 Perform the first structural projection. Calculate its eigenvalue decomposition, replace all negative eigenvalues ​​with small non-negative numbers (e.g., 0), and reconstruct the matrix; make the reconstruction result conform to the physical nature of the covariance matrix, and obtain the intermediate term J. _psd For J _psd Perform singular value decomposition (SVD) or eigenvalue decomposition. Based on the rank upper bound K set in the prior structure `Prior_struct`. _max Only keep the largest K _max Set one singular value (or eigenvalue) and the rest to zero, then reconstruct the matrix; this is used to enforce the sparsity of the model, resulting in the intermediate matrix J. _lr .

[0074] Furthermore, regarding J _lr Mean smoothing is performed along each diagonal direction to approximate the Toeplitz structure. A two-dimensional Fourier transform is then applied to the result, zeroing or attenuating spectral components beyond the bandwidth determined by the physical aperture of the light source, achieving band-limited projection. After structural projection is completed, observation consistency re-projection must be performed, forcibly resetting the values ​​of the matrix elements observed in the sampling layout Plan to the actual measured values ​​in the sparse angularly coherent observations y, obtaining the intermediate matrix J. _cons After a single step or a finite number of iterations, the final J will be obtained. _cons J, as the angular coherence matrix of this period # Output. Further, based on the final observation residuals ||P _Ω (J _cons )-y|| and structural projection residuals (such as ||J) _cons -J _lr ||), estimate the angular domain uncertainty U # In this method, the parameters between each projection step are fixed, and there is no need for online adaptive adjustment, making it simpler and more direct to implement.

[0075] This invention designs a sequentially coupled projection reconstruction method with adaptive structural parameters, solving the problem of decreased reconstruction accuracy in existing technologies when applying structural priors, due to fixed prior parameters or mismatches with the current light field state. In the sampling phase, this method actively collects information valuable to the model itself, prior to maximizing the identifiability of structural parameters (such as the rank of the matrix, Toeplitz stationary scale, etc.). In other words, it actively collects information highly correlated with the model. In the reconstruction phase, multiple structural prior projections are organized into interdependent chains. Parameters dynamically resolved during projection steps (such as low-rank projection) (such as the estimated rank K) are used to construct the reconstruction. _estThis information is used to adaptively adjust the constraints (such as the smoothing window size) of subsequent projection steps (such as near-Toeplitz projection). This unidirectional flow and coupling of information in the projection chain breaks the isolation between priors, allowing the structural prior model to self-correct and adapt in real time based on current observation data. In dynamically changing lighting scenarios, the reconstruction algorithm can automatically track changes in the coherent structure of the light field, avoiding model mismatch errors caused by using outdated or inaccurate prior parameters, and achieving stronger adaptability and higher reconstruction accuracy than fixed-prior methods.

[0076] Example 5 describes an optional technical solution for an online reconstruction method based on minimizing the upper bound of reconstruction error. This method couples the sampling and reconstruction stages together and achieves higher-quality reconstruction through a unified, task-oriented optimization framework.

[0077] According to this embodiment, online measurement is used to obtain sparse angularly coherent observations, and reconstruction is used to obtain the angularly coherent matrix and angular domain uncertainty of the current period. This includes: planning the sampling layout and baseline selection based on the angularly coherent matrix of the previous period, with minimizing the upper bound of the single-step reconstruction error as the task-oriented criterion; performing online measurement to capture sparse angularly coherent observations based on the sampling layout and baseline selection; constructing and solving a unified optimization problem encompassing observation consistency, nuclear norm regularization, and multiple structural prior indicator functions; and resolving the angularly coherent matrix and angular domain uncertainty of the current period from the sparse angularly coherent observations through single-step coupled near-end mapping update.

[0078] Accordingly, the sampling layout and baseline selection are planned based on minimizing the upper bound of the single-step reconstruction error as the task-oriented criterion; whereby the task clearly points to the subsequent reconstruction steps, that is, the sampling is used to make the reconstruction result J # With the true matrix J _true The error between ||J # -J _true || As small as possible; it does not rely on the identification of intermediate parameters (such as rank), but directly optimizes the final reconstruction quality.

[0079] Furthermore, the sampling layout and baseline selection are planned, including: for any unsampled candidate angle pair in the previous period's angular coherence matrix, calculating its marginal contribution to reducing the upper bound of single-step reconstruction error, and constructing the upper bound contribution of the angle pair-level error; based on the ratio of the upper bound contribution of the angle pair-level error to the estimated sampling cost, combined with the observation matrix condition number constraint and the angular domain coverage balance constraint, selection is made to determine the sampling layout and baseline selection; specifically including:

[0080] Based on the mathematical properties of the reconstruction algorithm (such as the coupled proximal mapping below) and the upper periodic angular mutual coherence matrix J _prevFrom this information, the system can deduce the reconstruction error norm ||J # -J _true || _F The computable theoretical bound is given by , where ||.|| _F Let be the Frobenius norm; it is a function of the selected set of sampling points. For each candidate corner pair that has not yet been sampled, the system calculates how much the aforementioned upper bound of error will decrease if the corner pair is added to the sampling set; this decrease is the marginal contribution of the corner pair. The marginal contributions of all corner pairs together constitute the corner pair-level error upper bound contribution map, Bound_map. The system employs a greedy algorithm or combinatorial optimization strategy to maximize the total marginal contribution / total sampling cost, while simultaneously satisfying the observation matrix condition number constraint (to keep the values ​​stable) and the corner domain coverage balance constraint (to avoid excessive concentration of sampling points). From the contribution map, the optimal set of corner pairs is selected to form the final sampling layout and baseline selection plan.

[0081] Building upon this, after sparse angularly coherent observations y, a unified optimization problem is constructed and solved, encompassing observation consistency, nuclear norm regularization, and multiple structural prior indicator functions. The unified optimization problem includes a weighted observation consistency term, where the weight W of this term is... _obs The weighted consistency of observations is determined based on the error reduction benefits derived from sampling layout and baseline selection, as well as the quality labels associated with sparse angularly interfering observations. In other words, the observation consistency involved in the unified optimization problem is weighted, and it is determined by analyzing, sorting out, and determining the error reduction benefits derived from sampling layout and baseline selection, as well as the quality labels associated with sparse angularly interfering observations. It should be understood that observations evaluated as having high marginal contributions during the sampling phase and labeled as having good quality during the acquisition phase will be given higher weights during reconstruction.

[0082] Find a matrix J that simultaneously satisfies data consistency, low-rank property, and other physical priors; its mathematical form can be expressed as: arg min _J [0.5*||W _obs *(P _Ω (J)-y)|| F 2 +λ _nuc *||J|| _* +ι _psd (J)+ι _Toeplitz (J)+ι _Bandlimit (J)]; where the weighted observation consistency term is 0.5*||W _obs *(P _Ω (J)-y)|| F 2 The reconstruction result J is required to be at sampling point P. _ΩThe weighted error between the observed value and the value y is minimized. The nuclear norm regularization term, λ... _nuc *||J|| _* Using the nuclear norm ||J|| _* (i.e., the sum of the singular values ​​of the matrix) serves as a convex approximation of the matrix rank, and minimizing it encourages low-rank solutions. The structural prior indicator function, ι _psd (J), ι _Toeplitz (J), ι _Bandlimit (J) is a set indicator function. For example, ι _psd (J) takes the value of 0 when J is a positive semi-definite matrix, otherwise it is infinite; physical priors are imposed on the solution as hard constraints.

[0083] In this embodiment, a single-step coupled proximal mapping update method is used to solve this non-smooth convex optimization problem. Accordingly, the single-step coupled proximal mapping update includes: sharing a dual variable to associate and jointly process at least two structural prior indicator functions, achieving non-surface-level coupled solution among multiple structural priors. Alternatively, the implementation of the single-step coupled proximal mapping update, through the sharing of a dual variable, associates and jointly processes at least two structural prior indicator functions for non-surface-level coupled solution of multiple structural priors.

[0084] Accordingly, the system employs methods such as the Alternating Directional Multiplier Method (ADMM) or the Primitive-Dual Split Algorithm to decompose the original problem into multiple subproblems. These subproblems (e.g., one handling positive semidefinite constraints and another handling near-Toeplitz constraints) are not executed sequentially but are processed in parallel during iterations, exchanging information and coordinating through one or more shared dual variables. These dual variables enable the various independent constraints to reach a common optimal solution, achieving non-surface coupling between priors and avoiding the order dependency and error accumulation problems that may arise from sequential processing. The entire process converges after a finite number of iterations, yielding the periodic angular coherence matrix J. # And the angular domain uncertainty U, which can be derived from information such as dual residuals. # .

[0085] Alternatively, nuclear norm regularization can be replaced by other non-convex rank approximation functions (such as truncated nuclear norms or Schatten-p norms), which can yield more accurate low-rank solutions, but usually increase computational complexity. Regarding solution algorithms, in addition to ADMM, other advanced first-order optimization algorithms can be used, such as the primordial-dual hybrid gradient method (PDHG), which can adapt to different problem structures. Weight matrix W _obs It can also make adaptive adjustments during the iteration process to further improve reconstruction performance.

[0086] Another possible implementation is a structure-adaptive sampling plan (task-oriented minimization of the upper bound of reconstruction error), which includes: reading the line-of-sight cone Cone, the structure prior Prior_struct, and the upper-period angular coherence matrix J. _prev Based on a single-step reconstruction error ||J, the single-step reconstruction error is reduced by a single-coupling proximal mapping. # -J * || _F An approximation of the computable upper bound (using the Lipschitz constant and the observation condition number as parameters) is performed to obtain the corner-pair level error upper bound contribution Bound_map. The corner-pair level error upper bound contribution Bound_map, the channel geometry / aperture and timing budget from the structural prior Prior_struct, and the line-of-sight cone Cone are read. Channels are sorted using the upper bound descent / cost ratio, and observation matrix condition number constraints and corner domain coverage equalization constraints are added to obtain the candidate channel list Cand_list (with sorting and constraint flags). The candidate channel list Cand_list and the system clock budget are read, quotas and trigger timings are allocated, and the sampling layout and baseline selection Plan are output.

[0087] The sampling layout and baseline selection plan are read, amplitude and phase corrections are performed, and internal calibration parameters are generated. Data is acquired and normalized according to the plan to obtain sparse angularly interdependent observations y. _raw For y _raw Perform noise estimation and packet loss labeling, and output sparse angularly coherent observations y (with quality labels). Read the sparse angularly coherent observations y, sampling layout and baseline selection Plan, and structure prior_struct. Construct the observation operator P. _Ω The observation weight matrix W is generated based on the quality label and the upper bound of the descent gain for each corner pair in the Plan. _obs (High-yield, high-quality sample weights are large). Read the upper-period angular coherence matrix J. _prev Prior structure (Prior_struct), weighted observation operator (P) _Ω W _obs Calculate the following formula in one iteration (or with very few steps):

[0088] J # =arg min _J [0.5*||W _obs (P _Ω (J)-y)||2 2 +λ _nuc *||J||*+ι _psd (J)+ι _Toeplitz (J)+ι _Bandlimit (J)];

[0089] Among them ι _*For set indicator functions (violation of a set has an infinite cost), ||J|| _* For the nuclear norm. In implementation, splitting + ADMM or primal-dual proximal one-time update is used, sharing a dual variable associated with Toeplitz and PSD constraints, making the two non-surface-level coupled. The dual residual set Dual_res and the weighted observation operator (P) are read. _Ω W _obs ), the inter-angular coherence matrix J of this period # The observation residual, dual residual, and frequency band leakage are combined to form the angular domain confidence radius, and the angular domain uncertainty U is output. # (Give the upper bound of the ellipsoid radius or variance according to the angular domain).

[0090] This invention employs a collaborative mechanism of deep coupling between sampling and reconstruction, resolving the conflict between measurement completeness and system real-time performance. Furthermore, the goal of sparse sampling is defined, moving away from traditional blind coverage or indirect parameter identification, as minimizing the upper bound of single-step reconstruction error. This ensures that the marginal contribution of each sampling to the final reconstruction accuracy matches the investment of sampling resources, avoiding wasting measurement time budget in information-redundant or insensitive angular domains. In the reconstruction phase, a unified optimization framework of single-step coupled proximal mapping is used, employing shared dual variables to non-surface-level coupled solutions for multiple structural priors such as positive semi-definite, low-rank, and near-Toeplitz structures. Furthermore, the error upper bound reduction benefit calculated in the sampling phase is incorporated as a weight into the observation consistency term, allowing the decisions in the sampling phase to directly guide the solutions in the reconstruction phase, forming a complete and self-consistent closed loop. This enables the system to obtain high-fidelity angular coherence matrix estimates in a very short time (e.g., tens of milliseconds) using far fewer samples than traditional methods, achieving superior reconstruction accuracy and efficiency compared to existing technologies, providing a data foundation for subsequent anti-glare control.

[0091] Example 6 provides an optional implementation of a chance-constrained weight optimization method, which explicitly and probabilistically incorporates the uncertainty of the reconstruction result into the constraints of the optimization model to obtain a more statistically reliable control strategy.

[0092] According to this embodiment, the optimization of generating coherent mode weights before execution specifically includes: expressing at least one key performance constraint as a chance constraint satisfying a preset probability; using statistical information derived from angular domain uncertainty, transforming the chance constraint into a deterministic optimization constraint, and solving for the coherent mode weights before execution. The implementation flow of this method is as follows:

[0093] Before performing this optimization step, the system has already obtained the current period's angular coherence matrix J through a reconstruction step (as in Example 4 or 5). # And the corresponding angular domain uncertainty U # U# J was quantified # The possible deviation between the estimated and true values ​​of each corner location. In this embodiment, U # This can be further interpreted as statistical information about the reconstruction error; for example, J can be derived from it. # Each element or key performance indicator calculated from it (such as groove depth D) _notch The mean and variance of ).

[0094] For one or more performance constraints in a system, this method transforms them from traditional deterministic forms into probabilistic forms. For example, for anti-glare requirements, i.e., the groove depth D... _notch It must be no less than a certain threshold D _thresh Traditional deterministic constraint writing: D _notch (w _new )≥D _thresh ; Due to D _notch It is based on the uncertainty of J # The calculated result is itself a random variable. Therefore, this embodiment rephrases it as a chance constraint:

[0095] P{D _notch (w _new )≥D _thresh}≥1-ε _notch ;

[0096] Where P{...} represents the probability, 1-ε _notch It is the preset confidence level (e.g., 95%), while ε _notch This represents the permissible risk of violation (e.g., 5%); the calculated weight w under the influence of current uncertainty. _new The probability of achieving the required groove depth must be at least 95%. Similarly, other key engineering parameters, such as illuminance and flicker, can also be expressed as similar chance constraints.

[0097] Since chance constraints cannot be directly handled by standard optimization solvers, statistical information derived from angular domain uncertainty is used to transform them into deterministic optimization constraints. This transformation is typically achieved using probability inequalities. For example, using Chebyshev's or Cantelli's inequalities, an upper bound on the probability of a random variable deviating from its mean within a certain range can be given based on the mean and variance of the random variable. Assuming that the groove depth D is obtained through uncertainty propagation... _notch The predicted mean μ _D (w _new ) and prediction variance σ D 2 (w _newTherefore, the above chance constraint can be transformed into a (potentially nonlinear) deterministic constraint, with the general form being: μ _D (w _new )-k(ε _notch )*σ _D (w _new )≥D _thresh ; where k(ε) _notch ) is related to risk ε _notch The relevant safety factor, the larger the k value, the more conservative the requirement; the new deterministic constraint is to subtract the safety margin determined by uncertainty (variance) and risk preference from the predicted mean, requiring the adjusted performance index to still meet the threshold.

[0098] After converting all chance constraints into deterministic constraints, the corresponding ordinary constraints in the original optimization model are replaced. The objective function of the optimization problem can remain unchanged, but its feasible region is reduced due to the introduction of a more conservative safety margin. Based on this, the system solves the new deterministic optimization problem to obtain the optimal modal weight increment Δw, and generates the coherent modal weights w before execution. _exec In summary, this embodiment provides an advanced control method that enables risk-controlled decisions when faced with unavoidable measurement and reconstruction uncertainties, thereby improving the robustness and reliability of the system.

[0099] Optionally, when transforming opportunity constraints, if stronger assumptions can be made about the distribution of uncertainty (e.g., assuming it follows a Gaussian distribution), a more precise transformation method can be used to obtain more compact (i.e., less conservative) deterministic constraints than those based on Chebyshev's inequality, achieving better performance while ensuring the same probability reliability. Correspondingly, different constraints in the system can be assigned different risk levels ε; for example, flicker constraints directly related to human eye safety can be assigned extremely low risk ε, while non-critical indicators such as energy efficiency can be assigned relatively high risk ε.

[0100] As one possible implementation method, this embodiment can also adopt the following scheme:

[0101] Read the angular complex coherence amplitude profile Γ # With the target constraint set Target _set In the discrete set Θ of the view cone Cone _cone Define weight ρ _cone A linearization approximation is used to map the modal weight increment Δw to the predicted coherence Γ. _pred , forming L _notch (Δw)=∑ θ∈Θ_cone [ρ _cone (θ) ∣Γ _pred (θ)∣];To obtain the groove cost structure L_notch Read the constraint mapping Map _cons Engineering performance boundary Spec _limit The cost of deviations in calculations based on piecewise linear or convex approximations, color rendering index or TM-30 index, flicker SVM, and energy efficiency: L _spec (Δw)=∑ _k β _k penalty(metric _k (Δw;Map _cons )-spec _k ), where β _k The index weights are k, where k is the index number; the output engineering cost structure L is... _spec Read the target constraint set Target _set Feasible domain _set The modal weight increment Δw is mapped to the expected execution time T using the actuator rate and amplitude boundaries. _pred Construct L _Dyn (Δw)=γ*max (0, T) _pred (Δw)-T _resp_max ); obtain the dynamic cost structure L _Dyn .

[0102] Reading the uncertainty weighted profile Gw # Write the groove and the engineering boundary as probability constraints and convert them into deterministic margins (based on Chebyshev or Cantelli bounds): Probability constraint illustration: P{D _notch (Δw)≥D _thresh}≥1-ε _notch ;P{metric _k (Δw)≤spec _k}≥1-ε _k The angular domain variance is derived from the uncertainty-weighted profile Gw. # propagation to Γ _pred Deviation from the indicator is used to obtain a safety margin and incorporate it into the constraint. A risk sensitivity cost L is constructed. _risk (Δw)=∑ _θ ρ _risk (θ) Var(Γ _pred (θ)); Output uncertainty constraint and risk cost L _risk (Including deterministic constraints and cost terms).

[0103] Integrated Groove Cost Structure L _notch Engineering cost structure L _spec Dynamic cost structure L _Dyn Uncertainty constraints and risk costs L _riskIncorporating hard constraints: Hard constraint 1 (groove depth threshold): D _notch (Δw)≥D _thresh (From the target constraint set Target) _set Hard constraint 2 (feasible region): Δw ∈ feasible region. _set (Actuator speed and amplitude limits). Objective function: min _Δw L _notch (Δw)+λ _1 L _spec (Δw)+λ _2 L _Dyn (Δw)+λ _3 L _risk (Δw); Choose a small-scale quadratic programming method or a projection iterative method to solve for the coherent mode weights w in this period. _new =w _prev +Δw.

[0104] This invention introduces an optimization method based on chance constraints, providing an alternative path to address model uncertainty and achieving probabilistic guarantees for key performance indicators. This method transforms rigid requirements such as anti-glare groove depth not being less than a threshold into chance constraints where this requirement holds with a probability of at least 95% (or other preset values). This allows decision-makers to intuitively set acceptable risk levels (e.g., a 5% failure probability) based on the safety level of the application scenario. Furthermore, this method utilizes statistical information derived from angular domain uncertainty (such as mean and variance) and mathematical tools like Chebyshev's inequality to transform probabilistic chance constraints into deterministic optimization constraints with safety margins. In practical applications of anti-glare lighting, when planning control strategies, the system not only considers performance based on the current best estimate but also reserves a safety boundary determined by the current uncertainty magnitude and acceptable risk level. Therefore, even if the reconstructed coherence matrix contains errors, the final control commands can still ensure the anti-glare effect meets the standards; this improves the system's reliability in the face of measurement noise and model errors, transforming uncertainty from a disruptive factor into a system parameter that can be quantified, managed, and controlled.

[0105] Example 7 provides a specific implementation process of a set-based robust weight optimization method. It adopts a deterministic set-based robust optimization method based on worst-case analysis, so that the system performance still meets the requirements under the most unfavorable perturbation.

[0106] According to this embodiment, the optimized generation of pre-execution coherent mode weights specifically includes: constructing a set of deterministic uncertainty domains that defines the perturbation range of the reconstruction result based on the angular domain uncertainty of the current period, without introducing probability distribution assumptions; and, on this set of deterministic uncertainty domains, calling and solving a set-based robust optimization problem to ensure that key performance constraints are still satisfied under the worst-case perturbation within the set, thereby obtaining the pre-execution coherent mode weights. Specifically:

[0107] The system will reduce the angular domain uncertainty U # This can be interpreted from statistics (such as variance) as geometric boundaries. For example, U can be... # Transformed into a constrained real angular coherence matrix J _true With estimation matrix J # Uncertain set U of the deviation range between _set This set does not rely on any assumptions about the error distribution (such as a Gaussian distribution), and is therefore more general and robust. Uncertain set U _set It can take many forms, for example, norm sphere constraint: U _set ={J|||JJ # || _F ≤R}, where R is based on U # The radius calculated from the total energy, ||.|| _F Let J be the Frobenius norm, indicating that all possible real matrices lie in the form of the Frobenius norm. # Within a hypersphere with center R and radius R, element-level interval constraints: U _set ={J||J(i,j)-J # (i, j)|≤U # (i, j), ⊿(i, j)∈{1,…,m}×{1,…,n}}; where ⊿ represents that all combinations within the range are satisfied; an independent perturbation interval is defined for each element of the matrix, forming a hyperrectangular uncertainty set.

[0108] After defining U _set Then, the system constructs and solves a robust optimization problem on this set of deterministic uncertainties; the deterministic constraints in the original optimization problem are replaced with their robust counterparts. For example, the original constraint D... _notch (J # w _new )≥D _thresh Its robustness counterpart is: min _J∈U_set D _notch (J, w) _new )≥D _thresh This constraint requires that, for an uncertain set U _set All possible Js, even the one that would make the groove depth D _notch The constraint must hold for J to become the minimum in the worst case.

[0109] In this embodiment, the construction of the robust optimization problem further includes: allocating a preset total risk budget across multiple discrete angles within the line-of-sight cone to form an angle-domain risk budget distribution; introducing a set of slack variables associated with the angle-domain risk budget distribution, and using these slack variables to quantitatively weigh the performance margin under worst-case perturbations in the constraints or objective function of the optimization problem. It should be understood that this process transforms the robust constraints into a larger but tractable deterministic optimization problem: setting a total performance margin budget or risk budget B. _total The total budget is allocated to each discrete angle θ within the line-of-sight cone Cone, forming an angular risk budget distribution b(θ) that satisfies ∑(b(θ))=B _total The principle of allocation can be based on uncertainty U. # Higher angles allocate more budget. For each angle θ, a non-negative relaxation variable ξ(θ) is introduced for the groove depth constraint. The robust constraint is relaxed to: min _J∈U_set D _notch_θ (J, w) _new )≥D _thresh_θ -ξ(θ); allows for performance at angle θ to be lower than the original threshold ξ(θ) in the worst case. The sum of all slack variables is constrained by the total budget: ∑(ξ(θ))≤B _total Slack variables are added to the overall objective function of the optimization problem in a weighted manner as a penalty; for example, increasing λ. _risk *∑(w _risk (θ)*ξ(θ)), where w _risk (θ) is the weight associated with risk budget allocation.

[0110] Building upon this foundation, the system solves an extended convex optimization problem that incorporates modal weight increments Δw and slack variables ξ. Its solution, while ensuring robustness to bounded uncertainties, also provides a quantifiable trade-off between performance and robustness through a risk budget mechanism.

[0111] Another example is the uncertain set U. _set The geometry can be customized based on prior knowledge of the perturbation structure; for example, if the perturbation is known to have a specific orientation, an ellipsoidal uncertainty set can be used. Optionally, the risk budget allocation strategy can be dynamic, adjusted based on real-time performance feedback and user preferences. Robust optimization can be applied not only to constraints but also to the objective function, forming a min-max problem to minimize the worst-case cost function value.

[0112] In another specific embodiment, the general chance constraint and variance penalty are replaced with ensemble robustness + risk budget allocation (corner domain decomposition), including: reading the uncertainty weighted profile Gw. # Angular complex coherence amplitude profile Γ # Target constraint set _set The angular domain uncertainty U # This is converted to a set radius over an angular domain (e.g., each angle θ has an interval or ellipsoidal radius r(θ)). Without introducing probability assumptions, we obtain J. # With Γ # The possible perturbation set U _set Read the target constraint set Target _set (Including groove depth threshold and response window) and U _set The total allowed risk budget B _total (Engineering configuration) Assign b(θ) to discrete points in the angular domain such that ∑θb(θ) = B _total Allocation principle: b(θ) is inversely proportional to r(θ) and directly proportional to the weight of the edge of the viewpoint cone. Read the constraint mapping Map. _cons Engineering indicator boundaries Spec _limit Feasible domain _set Upper period coherent mode weight vector w _prev Formation Δw→Γ _pred The first-order approximation of the changes in each index is derived from Δw; the actuator rate and amplitude boundaries are written as a cone constraint set of Δw, resulting in the constraint set (hard constraints) and the linear approximation coefficient set. The angular domain risk budget distribution Budget_map and the uncertainty domain set U are then read. _set Linear approximation coefficient set, angular complex coherence amplitude profile Γ # Target constraint set _set Feasible domain _set .

[0113] Construct robust groove constraints for angular domain decomposition: for each angle θ, require that in J∈U _set Under all disturbances, D _notch (Δw;θ)≥D _thresh -ξ(θ), let the slack variable ξ(θ)≥0, and ∑θξ(θ)≤B _total (That is, instead of including the probability, the total relaxation budget is used to control the relaxation of the worst-case perturbation). Construct the objective function: min _Δw L _notch (Δw)+λ _1 L _spec (Δw)+λ _2 L _Dyn (Δw)+λ _3 ∑θw_risk (θ)ξ(θ); where w _risk (θ) is consistent with Budget_map. The risk is incorporated into the optimization as an angular domain slack variable, determined by the total budget B. _total Constraints avoid the problem of missing probabilistic models and explicitly reflect the risk-performance trade-off. The solution outputs the coherent mode weights w for the current period. _new =w _prev +Δw.

[0114] In some embodiments, a robust optimization method based on risk budget allocation is introduced to address the problems of unstable control performance and lack of guaranteed anti-glare effect caused by model uncertainty. This method abandons the traditional approach of making probabilistic distribution assumptions about reconstruction errors. Based on quantified angular domain indefiniteness, it directly constructs a set of deterministic uncertainties defining all possible perturbations. This approach requires no complex statistical inference, is more direct, and has stronger adaptability to unknown perturbations. Furthermore, this method does not seek an absolutely conservative solution under all constraints, but introduces the concepts of total risk budget and angular domain risk budget distribution. This allows for small, controlled relaxations in system performance under worst-case perturbations, but the sum of all relaxations must be limited to the total budget. This mechanism transforms abstract robustness into resources that can be quantified and flexibly allocated among different angles and performance indicators. In anti-glare lighting scenarios, this enables the controller to make smarter decisions. For example, less risk budget can be allocated to the central visual area to ensure absolute glare-free operation, while more budget can be allocated to relatively less important peripheral areas or energy efficiency indicators. This risk management model enables the system to avoid performance losses caused by overly conservative design, while ensuring that key performance indicators (such as groove depth) meet robustness requirements, thus achieving an optimal balance between control robustness and overall system performance.

[0115] Example 8 describes the specific processing flow for the generation, propagation, and application of angular domain uncertainty, providing a robust means for the basic optimization model.

[0116] In this invention, generating angular domain uncertainty specifically includes: quantifying the observation residual based on the deviation between sparse angularly coherent observations and the current-period angularly coherent matrix; quantifying the structural projection residual based on the deviation between the current-period angularly coherent matrix and the structural prior; and combining the observation residual and the structural projection residual to jointly determine a quantified uncertainty index distributed in the angular domain. Angular domain uncertainty U # Used for quantization reconstruction matrix J # The confidence level of each element or each corner region. Its generation process integrates information from two dimensions:

[0117] Accordingly, the quantification of the observation residuals and the system calculation of the reconstruction results J #The degree of agreement between the sampled location and the true observed value y. Specifically, the observation residual vector r is calculated. _obs =P _Ω (J # Since observations are sparse, the residual vector is also sparse. The system uses interpolation or model-based methods to extend the sparse residual energy distribution across the entire angular domain, forming an observation residual map. On this map, regions with larger values ​​indicate that the model has difficulty fitting the true data in those regions, thus resulting in higher uncertainty.

[0118] Furthermore, to quantify the structural projection residuals, the system will evaluate J. # To what extent do the pre-defined structural a priori conditions apply? For example, calculating J... # Its projection on the positive semidefinite cone Proj _psd (J # The distance between ||J # -Proj _psd (J # The distance reflects J. # The degree of deviation from positive semidefiniteness can be calculated. Similarly, the degree of deviation from low-rank, near-Toeplitz, and other priors can be calculated to form their respective structural projection residual plots. On the plot, areas with larger values ​​represent conflicts between the data and prior assumptions, also indicating higher uncertainty.

[0119] Based on this, the aforementioned observation residual map and multiple structural projection residual maps are weighted and fused, for example, U # =α*R _obs _map+β*R _psd _map+γ*R _lr _map+..., where α, β, γ are weighting coefficients. The resulting U # Is with J # For matrices of the same dimension or functions distributed over angular domains, the higher the value, the less reliable the reconstruction result at the corresponding location.

[0120] In this embodiment, the propagation and application of angular domain uncertainty generate U # The method further includes: deriving the angular complex coherence amplitude profile from the current period's angular mutual coherence matrix; propagating the angular domain uncertainty to the angular complex coherence amplitude profile to generate an uncertainty-weighted profile; and applying an increased cost weight to the angular domain with higher uncertainty by utilizing the uncertainty-weighted profile in the optimization generation of coherent mode weights before execution.

[0121] Specifically, the system from J # The angular complex coherence amplitude profile Γ, describing the coherence intensity at each angle, is calculated. # The system will J #Uncertainty (by U) # (Description) Propagation to Γ # Above; it can be achieved through error propagation theory (e.g., using the Jacobian matrix for first-order Taylor expansion approximation) or Monte Carlo simulation, etc.; after propagation, Γ is obtained. # Estimation of variance or standard deviation σ at each angle θ _Γ (θ). Variance profile σ Γ 2 (θ) is the uncertainty weighted profile Gw # The specific form.

[0122] Based on this, when constructing the objective function of the optimization problem, an uncertainty-weighted profile is used to apply increased cost weights to the angular domains with higher uncertainty; this can introduce an uncertainty penalty term L. _uncert To achieve this. For example, L _uncert (Δw)=∑ _θ [Gw # (θ)*|Γ _pred (θ)|]. In the cost term, Gw # (θ) is used as a weight so that at angles θ with high reconstruction uncertainty, the residual coherence of any prediction |Γ _pred (θ)| will be disproportionately amplified and penalized; this drives the optimizer to generate more conservative solutions, i.e., to prioritize suppressing the coherence of regions with low confidence and to make safer choices under uncertainty.

[0123] One alternative implementation is to generate U # Furthermore, other information sources can be considered, such as the local density of sampling points; regions with low sampling density naturally have higher uncertainty. During the uncertainty propagation phase, higher-order analytical methods can be employed to obtain more accurate variance estimates. In the application phase, Gw... # It can be used not only to add independent cost terms, but also to dynamically adjust the groove cost L. _notch The internal angle weight ρ(θ) enables deeper fusion.

[0124] Example 9: A specific implementation scheme for an online adaptive update method for constraint mapping; wherein, constraint mapping (Map _cons This refers to the mathematical relationship model from the coherent mode weight vector w to various engineering indicators (such as illuminance, color rendering index, flicker, etc.). The accuracy of this model directly determines the quality of the optimization solution. The mechanism in this embodiment enables the model to continuously track changes in system state and maintain high accuracy.

[0125] According to this embodiment, in the optimization of generating coherent mode weights before execution, the constraint mapping used to estimate engineering indicators is updated online adaptively; the update process specifically includes: based on the execution records of at least one previous cycle, iteratively correcting the approximate mapping relationship from coherent mode weights to at least one engineering indicator to generate an updated constraint mapping for use in the current cycle.

[0126] Furthermore, the relationship between the coherent mode weights w and the actual luminance and electrical properties is not constant; it is subject to slow drift due to various factors such as ambient temperature, aging of the light source, and driver nonlinearity. If a fixed, offline calibrated map is used... _cons Its prediction accuracy decreases over time, causing the optimizer to make decisions based on incorrect information, which may cause the actual output lighting effect to deviate from the engineering specification boundary. _limit The online update mechanism in this embodiment provides the system with the ability to compensate for drift, enhancing the system's long-term robustness and accuracy.

[0127] The specific online update process may include the following steps: at the end of each or every N control cycles, the system records data pairs (w _exec M _measured ), where w _exec It is the coherent mode weight vector that is finally issued and executed in this cycle, while M _measured This is a vector of engineering parameters actually measured by sensors (e.g., miniaturized spectrometers or photometers) connected to the lighting system; this data is stored in a sliding window or historical database with a fixed length (e.g., the most recent 100 data points). Accordingly, the update process can be triggered periodically (e.g., once per minute) or by events; when the system detects the current Map _cons When the residual between the predicted value and the actual measurement value of a certain indicator continuously exceeds a preset threshold, an update process is initiated. Once the update is triggered, the system will utilize data pairs accumulated in the historical database to update the Map. _cons The model parameters are then corrected.

[0128] In a preferred embodiment, the correction process follows a four-step method: small perturbation-regression-regularization-projection. First, a batch of small perturbation data points adjacent to the current working point (i.e., the current w) is selected from the historical database to maintain local accuracy in the update. Then, the selected data pairs are used as training samples for Map... _cons The model parameters are refitted; for example, if Map _consIf the model is linear, M≈A*w+b, then ridge regression can be used to solve for the new parameter matrix A and bias b. The L2 regularization term introduced by ridge regression prevents overfitting when data is insufficient or noisy, enhancing the stability of the update. The new model parameters obtained through mathematical regression may not conform to physical laws (e.g., increasing a certain weight cannot lead to an infinite improvement in energy efficiency); therefore, the newly obtained parameters can be projected onto a set of feasible parameters defined by prior physical knowledge for correction. The updated map is then used... _cons_new Predict historical data and examine the residuals; if the overall prediction error of the new model is smaller than that of the old model, accept the update and Map. _cons =Map _cons_new Otherwise, abandon this update and may adjust the regression algorithm parameters (such as regularization strength) in the next update. From the above, Map... _cons It can continuously learn and evolve, keeping its description of the system's physical behavior fresh and accurate, thus providing a foundation for higher-level optimization decisions.

[0129] Another alternative implementation is to also map _cons Online learning algorithms employing nonlinear models, such as multinomial models, radial basis function networks, or small artificial neural networks. For example, stochastic gradient descent (SGD) updates based on mini-batch data can be used for neural network updates. Furthermore, during periods of low activity or permitted intervals in the lighting task, the system can proactively apply small, specially designed perturbations to w to collect data for updating the map. _cons This information is also known as active system identification.

[0130] Example 10 describes an optional implementation of the energy-equivalent actuator instruction mapping method, and elaborates on the optimized abstract pre-execution coherent mode weights w. _exec The process of converting the phase state into a sequence of actuator instructions u that can be executed by physical hardware. This process follows the principle of energy equivalence and is used to achieve seamless and imperceptible coherent state switching.

[0131] According to this embodiment, the pre-execution coherent mode weights are mapped to the actuator instruction sequence, following the principle of energy equivalence. The principle requires that the generated actuator instruction sequence, while realizing the target coherent state defined by the pre-execution coherent mode weights, maintains the corresponding illumination intensity and spectral characteristics with minimal perturbation compared to before execution.

[0132] The implementation of this step relies on a pre-established calibration map (Calib_map). This map describes the correspondence between the actuator instruction sequence u (e.g., the phase map grayscale values ​​of an SLM, the timing sequence of a DMD, the pump currents of each laser, etc.) and the coherent mode weight vector w. It typically has the following characteristics: First, it is non-unique (multiple solutions), meaning there are usually multiple sets of different actuator instructions u that can produce the same or very similar w. This provides degrees of freedom for selecting the optimal instruction. Second, it is complex, as this map is usually highly nonlinear and difficult to express analytically.

[0133] Specifically, the process of acquiring the Calib_map can include: systematically scanning the actuator instruction space in a controlled factory environment, measuring the corresponding output light field using high-precision optical equipment, and obtaining a large number of (u, w) data pairs. Based on this, a mathematical model can be fitted to represent the Calib_map, such as a high-dimensional interpolation table, a polynomial response surface model, or a trained neural network model. During system operation, the local parameters of the Calib_map can be fine-tuned by applying small perturbations and observing the response to compensate for factors such as temperature drift.

[0134] With Calib_map, the pre-execution coherent mode weights are mapped to the actuator instruction sequence. Specifically, this includes: using a pre-established calibration mapping, the pre-execution coherent mode weights are converted into actuator instructions; the conversion process follows the energy equivalence principle to ensure that the illumination intensity and spectral characteristics corresponding to the set of actuator instructions are kept to a minimum compared to before execution.

[0135] Accordingly, given the target w _exec Find u such that Calib_map(u)≈w _exec Due to the non-uniqueness of the solution, in all cases where Calib_map(u)≈w _exec Among the candidate instructions u, the solution that minimizes the perturbation of light intensity and spectral characteristics is selected. This can be formalized as a constrained optimization problem:

[0136] min _u ||P(u)-P(u _prev )|| 2 ;st||Calib_map(u)-w _exec || 2 ≤ε _tol ;

[0137] Among them, u _prevThis is the executor instruction from the previous cycle. P(u) is a function that calculates key photometric / radiometric parameters based on instruction u, such as total luminous flux (corresponding to illuminance) and color coordinates (corresponding to spectral characteristics); this function can also be part of the offline calibration. ε _tol It is a small tolerance, indicating how much deviation is allowed between the final achieved modal weights and the target values.

[0138] By solving this optimization problem, the system can find a new executor instruction u. _new This set of instructions accurately reproduces the coherent state of the target. _exec While achieving anti-glare, it maintains the stability of total brightness and color to the greatest extent, avoiding any flickering or color jumps that are perceptible to the human eye during adjustment, thus improving the user experience.

[0139] Optionally, during offline calibration of Calib_map, its null space is specifically sought and parameterized, meaning the direction of change of the photometric index P can be altered without changing w. During online mapping, a value satisfying Calib_map(u)≈w is found. _exec The particular solution is then adjusted along the pre-calculated null space direction to match the photometric index P(u) of the previous cycle as quickly as possible. _prev Furthermore, the scope of energy equivalence can be expanded. For example, in addition to illumination and spectrum, it is also possible to require minimizing disturbances to the total power consumption of the system, thereby achieving a more generalized, disturbance-free switching.

[0140] Example 11 provides a step-by-step reproducible numerical calculation case, taking a structure-prior-driven single-step projection reconstruction process as an example. Specifically, it includes:

[0141] Suppose a lighting system whose emission angles are discretized into N=4 angles. Therefore, its angular coherence matrix J is a 4×4 complex Hermitian matrix. Assume the true state of the light field at the current moment is an ideal rank-1 matrix: J _true ={{1.00, 0.71, 0.50, 0.35}, {0.71, 0.50, 0.35, 0.25}, {0.50, 0.35, 0.25, 0.18}, {0.35, 0.25, 0.18, 0.13}} (For simplicity, only the real part is shown here); Assume the state of the previous cycle was: J _prev ={{0.90, 0.60, 0.40, 0.20}, {0.60, 0.55, 0.30, 0.15}, {0.40, 0.30, 0.30, 0.10}, {0.20, 0.15, 0.10, 0.10}}; Set Rank _max =1; the semidefinite minimum eigenvalue τ _psd =0.

[0142] This period involves sparse sampling at 3 locations to obtain observations (ignoring noise): the observation value at J(1,1) is y. _1 =1.00; the observed value of J(2,4) is y. _2 =0.25; the observed value of J(4,2) is y _3 =0.25 (Hermitian characteristic);

[0143] Initial value setting, iterative initial value J _0 =J _prev Projecting J onto a positive semidefinite set for computation _0 The eigenvalues ​​are: {1.43, 0.28, 0.05, -0.06}. The existence of a negative eigenvalue -0.06 violates the positive semidefinite property. According to τ... _psd Given the prior value of 0, we trim it to 0. The new eigenvalues ​​are: {1.43, 0.28, 0.05, 0.00}. We then reconstruct the matrix to obtain J. _psd ={{0.91, 0.61, 0.41, 0.22}, {0.61, 0.54, 0.29, 0.14}, {0.41, 0.29, 0.31, 0.11}, {0.22, 0.14, 0.11, 0.09}}; Project onto a low-rank set to compute J. _psd The singular values ​​are: {1.48, 0.25, 0.04, 0.00}. According to Rank... _max Given the prior value = 1, only the largest singular value 1.48 is retained, and the rest are set to zero. The matrix is ​​reconstructed using the first singular value and its corresponding singular vector, resulting in a rank-1 J matrix. _lr J _lr ={{0.98, 0.68, 0.46, 0.29}, {0.68, 0.47, 0.32, 0.20}, {0.46, 0.32, 0.22, 0.14}, {0.29, 0.20, 0.14, 0.09}}; It can be seen that J _lr It has begun to structurally approximate J. _true Project onto the near-Toeplitz set (this step is omitted in this example for simplicity); observe consistency is back-projected using the J obtained in the previous step. _lr The values ​​at the three observed locations are forcibly replaced with the observed value y: J # (1, 1) = y _1 =1.00; J # (2, 4) = y _2 =0.25; J # (4, 2) = y _3 =0.25; the obtained reconstruction matrix J for this period is 0.25. # For: J #={{1.00, 0.68, 0.46, 0.29}, {0.68, 0.47, 0.32, 0.25}, {0.46, 0.32, 0.22, 0.14}, {0.29, 0.25, 0.14, 0.09}};

[0144] Compare the final reconstruction results J # And the real J _true J _true ={{1.00,0.71,0.50,0.35},{0.71,0.50,0.35,0.25},{0.50,0.35,0.25,0.18},{0.35,0.25,0.18,0.13}};J # ={{1.00, 0.68, 0.46, 0.29}, {0.68, 0.47, 0.32, 0.25}, {0.46, 0.32, 0.22, 0.14}, {0.29, 0.25, 0.14, 0.09}}; It can be seen that although there are only 3 discrete observation points, by utilizing structural priors (semi-definite, low rank), the reconstruction algorithm recovers a result that is very close to the true matrix in terms of overall structure. The overall error of the matrix ||J # -J _true ||Much smaller than the initial guess||J _prev -J _true The error of || proves the effectiveness and feasibility of the method of the present invention.

[0145] Example 12 provides a specific device structure for realizing an angular mutual coherence measurement device, which includes a light source system, a beam splitting sampling module, a fast switching module, and a detection array.

[0146] The device mainly consists of a fast optical path switching array based on microelectromechanical systems (MEMS); this array contains M×N independently controllable micromirror units, each of which can switch angles within microseconds. After the light emitted from the light source passes through the first beam splitter, 5% of the optical power is guided to the measurement optical path without affecting the main illumination function.

[0147] The measurement process specifically includes: at the beginning of each measurement cycle, the control system sends control signals to the MEMS array according to the sampling plan. For example, if it is necessary to measure the mutual coherence between angles θ1 and θ2, the system will configure two specific micromirror units M(i1, j1) and M(i2, j2) so that they reflect the beams of the corresponding angles to the same detector position.

[0148] The system employs time-division multiplexing technology to complete the cross-coherence measurement of multiple angle pairs within a 10-millisecond measurement window. Specifically, the measurement of each angle pair occupies 0.5 milliseconds, including 0.1 milliseconds for micromirror switching time and 0.4 milliseconds for signal integration time. For sparse sampling of 20 angle pairs, the total measurement time is 10 milliseconds.

[0149] During each 0.4 millisecond integration time, the detector collects the superimposed intensity I of the two beams of light. _total The system uses fast modulation technology to perform four sub-measurements within the integration time: I1 is obtained by turning on only the first beam, I2 is obtained by turning on only the second beam, and I is obtained by turning on both beams simultaneously. _plus The two beams of light are superimposed in opposite phase to obtain I. _minus Based on the four measurements, the complex mutual coherence γ can be calculated. 12 =(I _plus -I _minus ) / (4sqrt(I1I2)).

[0150] To further improve measurement speed, the system is equipped with eight independent detection channels, capable of simultaneously measuring the mutual coherence of eight pairs of angles. Combined with the rapid switching capability of the MEMS array, 160 pairs of angles can be measured within 10 milliseconds, meeting the requirements for sparse sampling.

[0151] Example 13 describes the specific implementation process of physical realization and control of coherent modes, and explains how the coherent mode weight vector can be used to regulate the coherent characteristics of the optical field through physical devices.

[0152] The system employs a coherent mode generator based on a multimode fiber bundle; the generator contains K single-mode fibers, each carrying a basic coherent mode; the output ends of the fibers are arranged in a specific geometric pattern to form a fiber array, and its far-field interference pattern constitutes a tunable partially coherent optical field.

[0153] Accordingly, each single-mode fiber is connected to an independent laser diode; controlling the drive current of each laser diode can adjust the optical power of the corresponding mode, which directly corresponds to each component of the coherent mode weight vector w. For example, w = [0.5, 0.3, 0.2, 0, ...] indicates that the first mode contributes 50% of the power, the second mode contributes 30%, the third mode contributes 20%, and the remaining modes are off.

[0154] A piezoelectric phase modulator is installed at the output end of each optical fiber. By applying different voltages, the relative phase relationship between each mode can be precisely controlled; this determines the far-field interference pattern and affects the mutual coherence between different angles.

[0155] When it is necessary to reduce coherence within a specific angular range, the system can increase the weight of higher-order modes, which have rapidly changing phase distributions within the target angular range. Furthermore, time-varying random phase modulation is applied to the modes, with a modulation frequency higher than the human eye response frequency but lower than the detector bandwidth. Based on this, the power allocation and phase modulation parameters of each mode are optimized to achieve low coherence in the time-averaged sense within the target angular range.

[0156] Optionally, during the system initialization phase, different combinations of drive currents [I1, I2, ..., I] are iterated. K ] and phase-modulated voltage combinations [V1, V2, ..., V K Using a measuring device, measure the angular coherence matrix J corresponding to each combination. Establish a lookup table or train a neural network model to realize the transformation from the target coherent mode weights w to the actuator instruction sequence u=[I1, V1, I2, V2, ..., I...]. K V K The mapping of ].

[0157] Furthermore, specific experimental data are provided for verification experiments to validate the anti-glare effect. The experimental setup in this embodiment includes: a coherent mode generator based on 32 single-mode optical fibers, each fiber connected to a 650nm laser diode with a maximum output power of 50mW; a MEMS micromirror array measurement system containing 64×64 micromirror units; a high-speed CMOS camera for monitoring the light field distribution; and a visual comfort evaluation system.

[0158] According to one aspect of this application, the experimental procedure and results are as follows: Without adaptive control, all 32 laser diodes operated at the same power without phase modulation. The measured angular coherence matrix showed an average coherence of 0.85 within a ±15 degree range, indicating high spatial coherence of the light field. At this time, five test subjects observed the light source from a distance of 2 meters and all reported significant glare, with an average subjective discomfort score of 7.8 (out of 10).

[0159] The system detects that the observer's line of sight is directly in front at 0 degrees, and sets the line of sight cone to a range of ±3 degrees. Through optimization algorithms, the system calculates new coherent mode weights: the weights of the first 8 low-order modes are reduced to 0.1-0.2, the weights of the 9th-24th mid-to-high-order modes are increased to 0.3-0.5, and a 10kHz random phase modulation is applied to the modes.

[0160] Under the new control parameters, measurements showed that the average cross-coherence within the visual cone decreased to 0.25, while the cross-coherence outside the cone remained above 0.75. Photometric measurements showed that the total luminous flux decreased by only 8%, and the color temperature shift was less than 50K. The subjective discomfort scores of the five test subjects decreased to 2.3, all indicating a significant reduction in glare.

[0161] The test subject's head swayed left and right at a frequency of 0.5 Hz, with an amplitude of ±30 degrees. The system tracked the changes in gaze in real time, with an average response delay of 85 milliseconds. Throughout the dynamic process, the mutual coherence within the gaze cone remained below 0.3, resulting in good subjective comfort.

[0162] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. An adaptive anti-glare lighting method, characterized in that, include: Based on the line-of-sight cone, the engineering index boundary, and the previous period's angular mutual coherence matrix and the previous period's coherent mode weight vector, sparse angular mutual coherence observations are obtained through online measurement. Based on the sparse angular mutual coherence observations and the angular mutual coherence matrix of the previous period, the angular mutual coherence matrix of the current period and the corresponding angular domain uncertainty are reconstructed. Based on the current period's angular coherence matrix, angular domain uncertainty, and the previous period's coherent mode weight vector, optimization is performed using the mode weight increment as the decision variable to obtain the coherent mode weights before execution. The coherent mode weights before execution are mapped to the executor instruction sequence and issued. At the same time, execution records are collected to update the previous cycle angular coherence matrix and the previous cycle coherent mode weight vector for the next cycle.

2. The method according to claim 1, characterized in that, Online measurements yield sparse angularly coherent observations, and reconstructions provide the current period's angularly coherent matrix and angular domain uncertainty, including: Based on the upper period angular mutual coherence matrix, and taking minimizing the upper bound of the single-step reconstruction error as the task-oriented criterion, the sampling layout and baseline selection are planned. Based on the sampling layout and baseline selection, online measurements are performed to capture sparse angularly interfering observations; A unified optimization problem encompassing observation consistency, nuclear norm regularization, and multiple structural prior indicator functions is constructed and solved. Through single-step coupled near-end mapping update, the angular coherence matrix and angular domain uncertainty of the current period are analyzed from sparse angular coherent observations.

3. The method according to claim 2, characterized in that, The sampling layout and baseline selection were planned, and further included: For any unsampled candidate angle pair in the upper period angle-direction mutual coherence matrix, calculate its marginal contribution to reducing the upper bound of the single-step reconstruction error, and construct the upper bound contribution of the angle pair level error. Based on the ratio of the upper bound contribution of the corner pair error to the estimated sampling cost, and combined with the condition number constraint of the observation matrix and the corner domain coverage balance constraint, the sampling layout and baseline selection are determined.

4. The method according to claim 2, characterized in that, The unified optimization problem includes a weighted observation consistency term, the weight of which is determined by the error upper bound reduction benefit derived from sampling layout and baseline selection, and the quality label attached to sparse angularly interfering observations.

5. The method according to claim 2, characterized in that, Single-step coupled proximal mapping update includes: By sharing a dual variable and associating and jointly processing at least two structural prior indicator functions, non-surface-level coupling solutions between multiple structural priors are achieved.

6. The method according to claim 1, characterized in that, Online measurements yield sparse angularly coherent observations, and reconstruction provides the current-period angularly coherent matrix and angular-domain uncertainty. Further possibilities include: Based on the criterion of maximizing the identifiability of the structural parameters contained in the prior structure, a sampling layout and baseline selection are generated. Based on the sampling layout and baseline selection, sparse angular mutual interference observations were collected; The angular coherence matrix of the current period is reconstructed by sequential coupled projection, where the parameters resolved in one structural prior projection are used to adaptively adjust the constraints of another structural prior projection.

7. The method according to claim 1, characterized in that, Optimize the generation of coherent mode weights before execution, specifically including: Based on the angular domain uncertainty of this period, a set of deterministic uncertainty domains is constructed to define the range of disturbances in the reconstruction results, without introducing probability distribution assumptions; On this set of deterministic uncertain domains, a set-based robust optimization problem is invoked and solved to ensure that key performance constraints are still satisfied under the worst-case perturbation within the set, thereby obtaining the coherent mode weights before execution.

8. The method according to claim 1, characterized in that, Optimizing the generation of coherent mode weights before execution also includes: At least one key performance constraint is expressed as a chance constraint that satisfies a preset probability. By utilizing the statistical information derived from the angular domain uncertainty, the chance constraints are transformed into deterministic optimization constraints, and the coherent mode weights before execution are obtained by solving the problem.

9. The method according to any one of claims 1, 7, or 8, characterized in that, The objective function to be minimized during the optimization process is composed of at least the following three cost weights: The groove cost used to penalize residual coherence within the line-of-sight cone, calculated based on the angular mutual coherence matrix of this period; Engineering index cost calculated based on engineering index boundaries, used to penalize photometric or electrical indices for deviating from their preset range; The dynamic response cost, calculated based on the target constraint set, is used to penalize the rate or magnitude of change of coherent mode weights.

10. The method according to claim 1, characterized in that, The pre-execution coherent mode weights are mapped to an actuator instruction sequence, specifically including: Using a pre-established calibration mapping, the coherent mode weights before execution are converted into actuator instructions; The conversion process follows the principle of energy equivalence, aiming to minimize the disturbance in the light intensity and spectral characteristics corresponding to the set of actuator commands compared to before execution.

Images