Naked eye 3D display system based on light field reconstruction algorithm

Abstract

The invention discloses a naked-eye 3D display system based on a light field reconstruction algorithm, particularly relates to the technical field of naked-eye 3D, and comprises a controlled scattering medium light field acquisition environment, scattering feature enhanced light field data analysis, a hierarchical depth optimization engine and dynamic wavefront modulation display control. According to the invention, through the controlled scattering medium environment design, a scattering signal which is regarded as interference in the traditional technology is converted into a depth characteristic carrier, the data acquisition quality in a complex optical environment is obviously improved, richer and more reliable basic data are provided for light field reconstruction, and the data acquisition efficiency is improved. An innovative hierarchical deep optimization engine is combined with a multi-scale confidence evaluation and discontinuous region processing method, so that the problem of edge blur in a shielding boundary and a geometric structure complex scene is effectively solved, and the geometric consistency and detail fidelity of a reconstructed light field are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of naked eye 3D, more particularly, the present application relates to a naked eye 3D display system based on a light field reconstruction algorithm. BACKGROUND

[0002] The light field reconstruction naked eye 3D display technology is an advanced display scheme for realizing stereoscopic vision by precisely controlling the direction of light propagation, and its core is to collect multi-dimensional light field information of an object and reconstruct a three-dimensional scene that can be directly perceived by the human eye. The operation process of the prior art usually includes three standardized steps: firstly, multi-view image data is collected by a camera array under a conventional lighting environment; secondly, a depth map is calculated based on stereo matching or deep learning; and finally, light deflection control is realized through a liquid crystal phase modulator or a microlens array.

[0003] However, the prior art has three key defects: first, environmental light interference and optical noise significantly reduce the signal-to-noise ratio of the collected data, limiting the depth information recovery accuracy; second, the traditional depth optimization method has insufficient processing capacity for occlusion boundaries and complex geometric structures, and the reconstructed light field has edge artifacts and detail loss; and third, the static phase modulation mode cannot adapt to dynamic changes in environmental parameters, resulting in a decline in display stability. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the present application provides a naked eye 3D display system based on a light field reconstruction algorithm, which solves the problems raised in the background art by the following scheme.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a naked eye 3D display system based on a light field reconstruction algorithm, comprising: S1: controlled scattering medium light field acquisition environment: composed of environmental parameter configuration, active light source coding and multi-view synchronous acquisition, by adjusting the physical properties of aerosol medium, combining with structured light coding scheme, and using a multi-camera array to realize time synchronization, a special light field data acquisition environment is constructed; S2: light field data analysis with enhanced scattering characteristics: including medium scattering modeling, frequency domain-spatial domain joint decoupling and nonlinear depth feature extraction, a light transmission physical model is established based on Mie scattering theory, frequency domain signal separation and reconstruction is realized through windowed Fourier transform, spatial domain features are processed in combination with non-local mean filtering, and finally cross-view geometric correlation features are extracted using a cascade neural network architecture; S3: hierarchical depth optimization engine: performing multi-scale confidence evaluation, non-continuous region optimization and physical constraint iteration, first constructing a Gaussian pyramid for multi-scale gradient confidence fusion, then using an improved Markov random field model to process depth discontinuous regions, and finally introducing light path reversibility constraints for iterative optimization; S4: Dynamic wavefront modulation display control: composed of phase distribution calculation, real-time aberration compensation and multi-plane fusion, based on depth field wavefront phase calculation and helmholtz decomposition, real-time monitoring and compensation of optical aberration through embedded wavefront sensor, finally realize multi-focal plane light field fusion by using angular spectrum diffraction theory.

[0006] Preferably, the environmental parameter configuration includes: filling the closed darkroom with silica suspension particles with a particle size range of 1-5 μm and a density controlled at 0.5-1.5 g / m³ as a controllable aerosol medium, the refractive index of the medium is set to 1.45 and allows a floating tolerance of ±0.05.

[0007] Preferably, the active light source coding includes: using an LED array as a light source component, the array is composed of 450 nm wavelength blue LED, 520 nm wavelength green LED and 630 nm wavelength red LED arranged in a 1:1:1 ratio, the independent coding sequence of the three wavelength light sources is synchronously switched by a programmable current driver at a frame rate of 120 fps, each coding period contains a 12-bit binary Golay code sequence, the projection lens group is equipped with a fisheye lens with f=8 mm to realize a 120-degree field of view angle coverage, the coding pattern has a spatial unique two-dimensional structured light stripe generated according to the De Bruijn sequence, the stripe minimum line width is 0.5 mm, the pattern switching is synchronized with the camera and exposure, the FPGA trigger circuit timing jitter is less than 1 μs, the LED driving current peak value is 5 A and is maintained at 80% duty cycle in the pulse width modulation mode, the coding data is preloaded into the light source controller buffer through the gigabit Ethernet interface, the LED junction temperature is monitored in real time during the implementation process and the cooling fan speed is dynamically adjusted to keep the temperature within the range of 45±2℃, at the same time, a closed-loop light intensity feedback mechanism is used to adjust the driving current through the photodiode sampling value, so that the projection brightness fluctuation does not exceed 5% of the nominal value.

[0008] Preferably, the multi-view synchronous acquisition comprises: using a 5x5 rectangular array of 25 global shutter CMOS industrial cameras, each camera having a resolution of 4096x2160 pixels and being equipped with a 16mm focal length fixed focus lens, the camera array baseline spacing being controlled to be 60mm, all cameras being synchronously triggered by an FPGA main control board, the trigger signal being connected to the GPIO trigger port of each camera in a star topology, the timing jitter being controlled to be within ±0.2μs, the exposure time being set to 1 / 240s and being an integer multiple of 2:1 of the refresh rate of 120fps of the active light source, the image data being transmitted to a preprocessing server through a CameraLink HS interface, each transmission link using a single-mode optical fiber to ensure that the total bandwidth of the 25 video streams is not less than 375Gbps, during the acquisition process, a non-uniformity correction algorithm is applied to eliminate the dark current differences of the sensors, the working temperature of each camera is monitored by an integrated temperature sensor, and the power of the thermoelectric cooler is dynamically adjusted to maintain the sensor temperature within the range of 30±0.5℃, before each acquisition, a high-precision ceramic calibration board is used for sub-pixel level optical distortion correction, the feature point spacing error of the calibration board is less than 1μm, and finally the raw data is stored in a multi-frame time sequence container file in 16-bit RAW format, and the timestamp with a precision of ns and the spatial pose matrix data are embedded in the file header.

[0009] Preferably, the medium scattering modeling comprises: first constructing a light intensity distribution calculation model based on Mie scattering theory, which obtains three core input parameters: the original light intensity data of the active light source in the multi-view acquisition system, the pre-calibrated physical property parameters of the aerosol medium including the particle size distribution range and the medium density of the suspended silica particles, and the scattering angle observation values recorded by the 25 synchronous cameras; in the model calculation stage, the mapping relationship between the incident light intensity and the scattered light intensity is established based on the Mie scattering principle: first, the Mie coefficient is calculated according to the real part 1.45 and the imaginary part 0.05 of the complex refractive index of the suspended particles, and then the upper limit of the order of the series expansion is determined by the size parameter of the scattering particles and the wavelength of the incident light; then the angular distribution function numerical solution is generated by a recursive algorithm, and finally the theoretical scattered light intensity distribution diagram under each camera view is output.

[0010] Preferably, the frequency-space domain joint decoupling comprises: first performing windowed Fourier transform processing on the original image acquired by each camera view, using an adjustable local analysis window to calculate the frequency components region by region in the spatial domain, and decomposing the image into components of different spatial frequencies; then performing feature separation operation in the frequency domain, identifying and separating the low-frequency components representing direct light and the high-frequency components carrying scattering features according to the scattering physical model, retaining 5% to 15% of the high-frequency components as depth information carriers by setting a dynamic energy threshold, and suppressing the remaining frequency components; The filtered high-frequency components are then converted back to spatial domain representation and subjected to improved non-local mean filtering process, which adaptively smooths the selected high-frequency components while preserving edge features by computing similarity weights between image patches.

[0011] Preferably, the nonlinear depth feature extraction comprises: firstly constructing a three-stage cascaded deep learning architecture, the first stage adopts a three-layer convolutional neural network to process the preprocessed multi-view image patches, wherein the first layer uses a 7x7 convolution kernel to extract large-scale local geometric features to generate a 128-channel feature map, the second layer compresses and reduces the dimension of the features to 64 channels through a 5x5 convolution kernel, and the third layer captures texture details using a 3x3 convolution kernel to output a 32-channel feature tensor, and all convolution layers are followed by LeakyReLU activation functions and batch normalization operations; then the second stage introduces a graph convolution network to establish inter-view correlation, taking 25 camera views as graph nodes and constructing an adjacency matrix according to camera pose parameters, specifically defining the edge weight between nodes as the reciprocal of the Euclidean distance of the camera projection center multiplied by the cosine value of the optical axis angle, and realizing cross-view information fusion by iteratively aggregating neighborhood node features through three graph convolution layers; finally, the third stage uses a time sequence long short-term memory network to process dynamic encoding light sequences, taking consecutive 12 frames of Gray code images of each view as time sequence input, designing a phase shift sensitive unit in the memory gating mechanism, which calculates the local area cross-correlation value of adjacent frames as the input of the gating signal, and connecting a fully convolutional decoding layer at the output end of the network to map the spatio-temporal features into the initial depth map, wherein the final output layer uses a 1x1 convolution kernel with a Sigmoid activation function to generate normalized depth prediction values.

[0012] Preferably, the multi-scale confidence assessment comprises: firstly constructing three depth map representations of different scales on the original depth map, realizing multi-scale expression through Gaussian blur and downsampling operation; then independently calculating the spatial gradient amplitude on each scale depth map, converting the gradient amplitude to local confidence measure, and the specific method is to take the negative exponential mapping of the square of the gradient amplitude and take the complement, wherein the proportional coefficient of the exponential function is a preset constant Control; Finally, the confidence measures at the three scales are geometrically weighted and fused, and the weight coefficients of each scale are pre-set according to actual application requirements, and the fusion result is used as the final confidence assessment value of each pixel position for subsequent optimization decision.

[0013] Preferably, the non-continuous region optimization comprises: firstly, constructing a Markov Random Field model based on the initial depth map, the model including two parts of data term and smooth term, wherein the data term uses a non-convex function to calculate the difference between the current depth estimation value and the initial depth estimation value, and the non-convex function is set as the 0.8 power form of the absolute value of the depth difference to enhance the robustness to abnormal values; the smooth term uses a truncated linear function to calculate the depth difference between adjacent pixels, and when the absolute value of the depth difference is less than a preset threshold, the actual difference value is taken, and when the threshold is exceeded, a fixed truncated value is taken; the overall energy function is expressed by the weighted sum of the data term and the smooth term, wherein the smooth term weight coefficient is controlled by a global parameter; for the detected occlusion boundary region, the smooth term weight coefficient is separately set to one tenth of the global weight; finally, a graph cut algorithm is used to minimize the energy function to obtain the optimized depth map.

[0014] Preferably, the physical constraint iteration comprises: establishing an optimization constraint mechanism based on the optical reversibility principle, and the specific process is: firstly, using the current depth estimation value, the two-dimensional depth map is inverted into three-dimensional space point cloud data through inverse projection operation; then, the three-dimensional space point cloud data is projected to each camera perspective coordinate system to generate a corresponding perspective synthetic depth map; then, a pixel-level difference matrix between the synthetic depth map and the actual observed depth map under all perspectives is calculated; a target function is constructed as the square sum of the difference matrix of all perspectives; finally, the Levenberg-Marquardt optimization algorithm is used to iteratively update the depth map estimation value, and the gradient descent adjustment is used to minimize the target function.

[0015] Preferably, the phase distribution calculation comprises: firstly, receiving the depth map data from the hierarchical depth optimization engine output, and then calculating the initial value of the wavefront phase of each pixel point based on the physical optical principle, which converts the depth value into the phase value according to the phase delay characteristics of the light wave propagation in the medium; Then, a two-dimensional gradient field calculation is performed on the phase field to respectively solve the phase partial derivatives in X and Y directions to construct a gradient vector field; then, the gradient field is reconstructed into a continuous phase distribution by solving the Poisson equation, and a multigrid iteration algorithm is used to maintain the convergence stability at the occlusion boundary; subsequently, the Helmholtz decomposition is performed on the reconstructed phase to separate the divergence component and the irrotational component, and only the irrotational gradient component is retained as the final modulation reference, and finally the continuous phase distribution map conforming to the wavefront propagation law is output to the aberration real-time compensation process.

[0016] Preferably, the aberration real-time compensation iteration comprises: firstly, deploying micro wavefront sensors at four edge positions of a display screen respectively to continuously monitor the phase distribution state of the output light field, and the sensors synchronously collect actual wavefront distortion data at a sampling frequency of 120Hz; then inputting the collected distortion data into a preset aberration mathematical model, the model uses fifteen standard orthogonal basis functions to describe common aberration types; subsequently, dynamically solving the corresponding coefficient combination of the current aberration state by using a least square optimization algorithm, and introducing a sliding window filtering mechanism to eliminate the influence of instantaneous measurement noise; then converting the solved aberration coefficients into corresponding phase compensation amounts, and superimposing them on the original phase distribution map to generate a corrected modulation signal.

[0017] Preferably, the multi-plane fusion construction comprises: firstly, dynamically determining the axial positions of three virtual focal planes according to the depth distribution range of the scene, and the three planes correspond to the typical focusing distances of the foreground, the middle ground and the background respectively; then independently performing diffraction propagation calculation on each virtual focal plane to generate the intermediate light field distribution of the corresponding depth through a preset light wave transmission model; then calculating adaptive fusion weights based on the statistical characteristics of the depth map, and the weights are dynamically adjusted by the entropy value of the depth distribution to ensure the preferential presentation of the rich detail area; finally, performing weighted superposition processing on the intermediate light fields of the three focal planes, and introducing a phase alignment correction mechanism in the superposition process to eliminate interference artifacts between different planes, and generating the final full-depth light field output conforming to the human eye visual characteristics.

[0018] Technical effects and advantages of the present application: 1、The present application converts the scattering signal which is regarded as an interference factor in traditional technology into a depth feature carrier by constructing a controlled scattering medium environment, uses the Mie scattering model combined with the frequency-space joint decoupling method to extract the implicit depth information by non-local mean filtering on the basis of retaining the high-frequency scattering components, significantly improves the data acquisition quality in complex optical environment, and this innovative processing mechanism enables the system to analyze effective three-dimensional features from optical noise that cannot be utilized in traditional technology, thereby providing more abundant basic data for subsequent processing. 2、For the depth optimization link, the present application proposes a hierarchical confidence evaluation and non-continuous region optimization mechanism, realizes adaptive processing of occlusion boundaries and depth discontinuous regions by combining multi-scale gradient confidence fusion and an improved Markov random field model, and combining non-convex function and truncated linear function combination optimization, which effectively solves the edge blur problem of traditional schemes in complex geometric structure scenes, and significantly improves the geometric consistency and detail fidelity of the reconstructed light field. 3. At the display control level, the application innovatively introduces dynamic wavefront compensation and multi-plane fusion mechanism, realizes aberration correction through real-time wavefront sensing and Zernike polynomial fitting, combines depth distribution entropy adaptive multi-plane weighted fusion algorithm, so that the system has dynamic response capability to environmental parameter changes. This technology breaks through the inherent limitations of static phase modulation, greatly improves the robustness and stability of the system in dynamic environment while maintaining display accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a schematic diagram of the overall structure of the application.

[0020] Figure 2 is a schematic diagram of the controlled scattering medium light field acquisition environment structure of the application.

[0021] Figure 3 is a schematic diagram of the light field data analysis structure of the application with enhanced scattering characteristics.

[0022] Figure 4 is a schematic diagram of the hierarchical depth optimization engine structure of the application. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0024] Reference Figures 1-4 The application shows a naked-eye 3D display system based on a light field reconstruction algorithm, comprising: S1: Controlled scattering medium light field acquisition environment: composed of environmental parameter configuration, active light source coding and multi-view synchronous acquisition, through adjusting the physical properties of aerosol medium, combining with structured light coding scheme, and using multi-camera array to realize time synchronization, a special light field data acquisition environment is constructed.

[0025] Environmental parameter configuration: fill the closed darkroom with silica suspension particles with a particle size range of 1 μm to 5 μm and a density controlled at 0.5 g / m³ to 1.5 g / m³ as controllable aerosol medium, the refractive index of the medium is set to 1.45 and a floating tolerance of ± 0.05 is allowed.

[0026] Active light source encoding: LED array is used as light source component, the array is composed of 450 nm wavelength blue LED, 520 nm wavelength green LED and 630 nm wavelength red LED, which are arranged in 1:1:1 ratio, the independent encoding sequence of three kinds of wavelength light sources is switched synchronously by programmable current driver at 120 fps frame rate, each encoding period contains 12-bit binary Golay code sequence, the projection lens group is equipped with f=8mm fisheye lens to realize 120 degree field of view angle coverage, the encoding pattern has spatial uniqueness according to the De Bruijn sequence to generate two-dimensional structured light stripe, the minimum stripe width is 0.5mm, the pattern switching is synchronized with the camera and exposure, the timing jitter of FPGA trigger circuit is less than 1μs, the peak value of LED driving current is 5A and is maintained at 80% duty cycle in pulse width modulation mode, the encoding data is preloaded into the light source controller buffer through the gigabit Ethernet interface, the LED junction temperature is monitored in real time during the implementation process, and the speed of the cooling fan is dynamically adjusted to make the temperature in the range of 45±2℃, at the same time, the closed-loop light intensity feedback mechanism is used to adjust the driving current through the sampling value of the photodiode, so that the projection brightness fluctuation does not exceed 5% of the nominal value.

[0027] Multi-view synchronous acquisition: 25 global shutter CMOS industrial cameras are arranged in a 5x5 rectangular array, each camera has a resolution of 4096x2160 pixels and is equipped with a 16mm focal length fixed focus lens, the camera array baseline spacing is controlled to be 60mm, all cameras are triggered synchronously through the FPGA main control board, the trigger signal is connected directly to the GPIO trigger port of each camera in a star topology, the timing jitter is controlled within ±0.2μs, the exposure time is set to 1 / 240s and is kept in an integer multiple relationship of 2:1 with the 120fps refresh rate of the active light source encoding, the image data is transmitted to the preprocessing server through the Camera Link HS interface, each transmission link uses a single-mode optical fiber to ensure that the total bandwidth of 25 video streams is not less than 375Gbps, during the acquisition process, the non-uniformity correction algorithm is applied in real time to eliminate the dark current difference of each sensor, at the same time, the working temperature of each camera is monitored through the integrated temperature sensor and the power of the thermoelectric cooler is dynamically adjusted to maintain the sensor temperature in the range of 30±0.5℃, before each acquisition, a high-precision ceramic calibration board is used for sub-pixel level optical distortion correction, the feature point spacing error of the calibration board is less than 1μm, finally the original data is stored in 16-bit RAW format as a multi-frame time sequence container file, and the timestamp with a precision of ns and the spatial pose matrix data are embedded in the file header.

[0028] The S1 provides the multi-view light field sequence data carrying scattering features for the S2, wherein the light scattering effect introduced by the aerosol environment provides a physical basis for depth analysis, and the synchronous parallax data collected by the multi-camera array constitutes the basic input source for naked-eye 3D reconstruction.

[0029] S2: Scattering feature enhanced light field data analysis: contains medium scattering modeling, joint frequency-spatial domain decoupling and nonlinear depth feature extraction, a physical model of light transmission is established based on Mie scattering theory, frequency domain signal separation and reconstruction is realized through windowed Fourier transform, spatial domain features are processed combined with non-local mean filtering, and finally cross-view geometric correlation features are extracted using a cascade neural network architecture.

[0030] Medium scattering modeling: first, a light intensity distribution calculation model is constructed based on Mie scattering theory, which obtains three core input parameters: the original light intensity data of the active light source in the multi-view acquisition system, the pre-labeled physical property parameters of the aerosol medium including the particle size distribution range and medium density of the suspended silica particles, and the scattering angle observation values recorded by 25 synchronous cameras; in the model calculation stage, the mapping relationship between incident light intensity and scattered light intensity is established based on Mie scattering principle: first, the Mie coefficient is calculated according to the real part 1.45 and the imaginary part 0.05 of the complex refractive index of the suspended particles, and then the upper limit of the order of series expansion is determined by the size parameter of the scattering particles and the wavelength of the incident light; then the numerical solution of the angular distribution function is generated by the recursive algorithm, and finally the theoretical scattering light intensity distribution diagram under each camera view is output. The model continuously receives real-time light source encoding sequences for dynamic correction, and transmits the calculation results to the subsequent joint frequency-spatial domain decoupling process as the physical basis for feature separation. In this way, the traditional scattering signal, which is considered as noise, is transformed into a carrier of depth information.

[0031] Let the initial intensity of the incident light be , the wavelength of the incident light be , corresponding to the 450nm / 520nm / 630nm of the LED light source, the distance from the observation point to the scattering particle be , and the scattering angle be , then the scattering light intensity received by the observation point is , : the order of series expansion, starting from 1 and increasing to infinity, , : Mie coefficient (complex number), calculated from the complex refractive index (real part 1.45, imaginary part 0.05) and size parameter of the scattering particle, : angular distribution function, describing the electric dipole scattering component, : angular distribution function, describing the magnetic dipole scattering component; the construction logic of this formula starts from the strict solution of Maxwell's equations for uniform spherical particles, and the scattering process is decomposed into the superposition of multiple polar oscillations through spherical harmonic function expansion: first, introduce the spherical wave diffusion factor , which embodies the inverse square law of light intensity with distance, ensuring energy conservation; the core series term serves as the spherical harmonic expansion weight coefficient, maintaining the electromagnetic field continuity boundary condition; where and The two Mie coefficients quantify the contribution strength of the electric multipoles and the magnetic multipoles respectively, and are determined by the complex refractive index of the particles (1.45+0.05i for the silica suspension in this example) and the size parameter; the angular distribution function and Coupling describes the directional characteristics of scattered light, Convert the spatial angle into a computable scalar; finally, perform a modular square operation The complex amplitude is converted into observable light intensity. Its physical significance lies in accurately describing the angular distribution of light energy in a single scattering event: when the particle size parameter When , it degenerates into the Rayleigh scattering model. When the typical Mie resonance characteristics (such as forward scattering enhancement and rainbow ring) are presented, the Mie coefficient in the formula , As a medium fingerprint, it is sensitive to the chemical composition of the suspended particles (such as the change in the real and imaginary parts of the refractive index of silica suspended particles); in this embodiment, the geometric characteristics of the object surface are detected by modulating the local scattering angle Influence Intensity distribution (e.g. depth transition edge triggering The intensity mutation at the spot makes the scattering signal, which is traditionally regarded as noise, transformed into a carrier of depth information, which is the physical basis for realizing the reverse utilization of the scattering environment.

[0032] Frequency domain-spatial domain joint decoupling: First, the original image captured by each camera angle is processed by window Fourier transform, and the frequency components are calculated region by region in the spatial domain using an adjustable local analysis window to decompose the image into components of different spatial frequencies; then, feature separation operation is performed in the frequency domain, and the low-frequency components representing direct illumination and the high-frequency components carrying scattering characteristics are identified and separated according to the scattering physical model. By setting the dynamic energy threshold, the high-frequency components of 5% to 15% of the total energy are retained as the depth information carrier, and the remaining frequency components are suppressed; suppose the original image is at the spatial position The light intensity at , then the window Fourier transform formula is , : Movable local analysis window function, : two-dimensional complex exponential kernel function, :spatial integral differential, :Four-dimensional output function; its core design idea is to break through the spatial limitations of the traditional global Fourier transform, and to achieve the joint representation of position and frequency by introducing a movable local analysis window function in the spatial domain. The specific construction logic is first to Apply window function To achieve local signal interception, the window usually uses a Gaussian function to maximize the center point weight to ensure spatial positioning accuracy; then the intercepted local signal With complex exponential kernel Perform inner product operation and traverse all pixel positions in the window through integration Complete the basis transformation from spatial domain to frequency domain; finally output the four-dimensional function Include specific locations The complete local spectrum information at Respectively represent the spatial wave number in the x / y direction; this transformation establishes a direct mapping relationship between the spatial position and the frequency domain component: the low-frequency component u,v values ​​are small Corresponding to large-scale continuous direct illumination features, high-frequency components u,v value is large It carries the microstructural information of the scattering medium and can adaptively match scattering characteristics of different scales by dynamically adjusting the window size, ultimately achieving the innovative goal of converting high-frequency scattering signals, traditionally regarded as noise, into an effective carrier of depth information. The filtered high-frequency components are then converted back to spatial representation and subjected to an improved non-local mean filtering process. This filtering process calculates the similarity weights between image blocks and performs adaptive smoothing on the selected high-frequency components while retaining edge features. Finally, an optimized image with both spatial structural integrity and frequency domain feature significance is generated. This output is directly transmitted to the subsequent nonlinear depth feature extraction process. The entire processing chain transforms the scattering components traditionally regarded as interference signals into effective depth information sources through the synergy of frequency domain analysis and spatial domain optimization. Assume that the original image is The pixel value at position , with the target pixel The local image block centered at , with the current traversal pixel The local image block centered at , then the non-local mean filter formula is , : Pixels after filtering The output value of : Normalization factor to ensure that the weights sum to 1, : Weighted Euclidean distance in high-dimensional space, : Filter control parameter (bandwidth) determines the sensitivity of similarity judgment; this formula is based on the principle of image self-similarity, and constructs a weighted average mechanism by systematically comparing the similarity of the target pixel neighborhood with the similarity of all pixel neighborhoods in the entire image: first, the target pixel coordinates Extract local image patches for the center As a feature template, synchronously traverse each pixel coordinate in the image And extract the corresponding image block , by calculating the weighted Euclidean distance between the two to quantify the block similarity, the smaller the distance value, the higher the similarity; then substitute the distance value into the exponential weight function Generate weight coefficients. The physical meaning of this function is to give exponentially growing weights to high-similarity areas, and to impose rapidly decaying small weights on low-similarity areas. Finally, all pixel values ​​are Weighted summation by weight and normalized by coefficient The core physical meaning of ensuring that the output value is within a reasonable dynamic range is to utilize the repeated structural characteristics of the image itself to intelligently distinguish real signals from random noise in the spatial domain, enhance effective features through collaborative filtering of similar areas, and suppress isolated noise points. It is particularly suitable for preserving structured depth features in scattered signals.

[0033] What needs to be specifically explained in this embodiment is that is the analysis center coordinate, which indicates the reference position of the current window in the image. It is the integral traversal coordinate, which represents the spatial position of any pixel on the image.

[0034] Nonlinear deep feature extraction: First, a three-level cascade deep learning architecture is constructed. The first level uses a three-layer convolutional neural network to process the preprocessed multi-view image blocks. The first layer uses a 7×7 convolution kernel to extract large-scale local geometric features to generate a 128-channel feature map. The second layer uses a 5×5 convolution kernel to compress the features and reduce the dimensionality to 64 channels. The third layer uses a 3×3 convolution kernel to capture texture details and output a 32-channel feature tensor. All convolution layers are followed by LeakyReLU activation function and batch normalization operation; then the second level introduces a graph convolution network to establish the correlation between perspectives, takes 25 camera perspectives as graph nodes and constructs an adjacency matrix based on the camera pose parameters. The weight of edges between volume-defined nodes is the inverse Euclidean distance from the camera projection center multiplied by the cosine of the angle between the optical axes. Three layers of graph convolutions iteratively aggregate features from neighboring nodes to achieve cross-view information fusion. Finally, the third layer uses a temporal long short-term memory network to process dynamic coded light sequences. Twelve consecutive Gray-coded frames from each view are used as a time series input. A memory gating mechanism is implemented using a phase-shift-sensitive unit that calculates the local cross-correlation between adjacent frames as a gating signal. The network output is connected to a fully convolutional decoding layer to map spatiotemporal features into an initial depth map. The final output layer uses a 1×1 convolution kernel with a sigmoid activation function to generate normalized depth predictions. The entire process optimizes network parameters through end-to-end training. The loss function uses a Huber loss based on the ground-truth and predicted depth values ​​combined with a gradient similarity constraint. This architecture innovatively integrates spatial convolution, graph structure modeling, and time series analysis, achieving the first simultaneous modeling of local geometric consistency, viewpoint topological correlation, and dynamic coded light propagation in light field analysis.

[0035] The S2 receives the original light field data of S1 and extracts the depth feature tensor to deliver to S3, which enhances the three-dimensional spatial representation of the transparent object by decoupling the scattering signal, and the perspective correlation feature is established to provide the core geometry for naked eye 3D parallax reconstruction.

[0036] S3: Hierarchical depth optimization engine: performs multi-scale confidence evaluation, non-continuous region optimization and physical constraint iteration, first constructs a Gaussian pyramid for multi-scale gradient confidence fusion, then uses an improved Markov random field model to process depth discontinuous regions, and finally introduces light path reversibility constraints for iterative optimization.

[0037] Multi-scale confidence evaluation: first, construct three different scale depth map representations on the original depth map, realize multi-scale expression through Gaussian blur and downsampling operation; then calculate the spatial gradient amplitude on each scale depth map independently, convert the gradient amplitude to local confidence measure, the specific method is to take the negative exponential mapping of the square of the gradient amplitude and take the supplement, where the proportional coefficient of the exponential function is a pre-set constant control; set in scale The depth map after Gaussian blur and downsampling is , the gradient vector of the depth map at is , then the gradient confidence , : scale parameter, representing the level of the Gaussian pyramid (the value of this embodiment is 1.2, 2.0, 3.5), controlling the multi-scale expression of the depth map, : the two-norm of the gradient vector (i.e. gradient amplitude), the calculation formula is , : sensitivity adjustment constant (pre-set positive value), control the sensitivity of confidence to gradient change (the smaller , the higher the sensitivity); the gradient confidence formula is based on the physical correlation between spatial gradient and confidence in depth estimation: first, the square of the gradient amplitude of the depth map at a certain scale is calculated to quantify the intensity of local depth change, then the gradient amplitude is nonlinearly mapped to the interval (0, 1] using the exponential function, and finally the confidence is inverted by 1 minus the mapping value. The physical meaning is to convert the spatial structure features of the depth map into reliability quantitative indicators: in the high gradient area of object boundary, the formula outputs high confidence close to 1, which conforms to the physical law that the edge depth is more reliable in optical observation; in the low gradient area of flat area, the formula outputs low confidence close to 0, reflecting the essential characteristics that these areas are susceptible to noise interference; and the square operation strengthens the contribution of significant edges, and the exponential function suppresses minor fluctuations, ultimately forming a confidence evaluation model that conforms to the physical law of depth perception, providing accurate reliability basis for multi-scale optimization; Finally, the confidence measures of the three scales are geometrically fused with the weight coefficients of each scale being preset according to the actual application requirements, and the fusion result is used as the final confidence evaluation value of each pixel position for subsequent optimization decision making; let the fused multi-scale confidence , : the final output pixel-level fusion confidence value, the value range is [0, 1], : the preset scale weight coefficient; the formula is based on the principle of cross-scale depth feature consistency: first, the gradient-driven confidence is independently calculated on the depth maps of different scales (small scale captures details, large scale represents structures), and then the exponential weighting operation is used to realize scale contribution regulation, when the weight is greater than 1, the key scale effect is amplified, when the weight is less than 1, the influence of the secondary scale is weakened; the final product operation constitutes a serial decision mechanism, which requires that the target feature must be stable in all scales to obtain a high value, and its physical meaning embodies three core characteristics: first, the multi-scale spatial response is converted into a probabilistic reliability measure, and tends to 1, indicating that the feature has cross-scale consistency; second, a noise immunity mechanism is established, and single-scale outliers will cause the overall confidence to collapse; third, the edge protection capability is strengthened, and the small-scale high confidence in the depth discontinuous area dominates the fusion result through exponential amplification, and this design realizes the balance between noise suppression and feature preservation through nonlinear cross-scale constraint, which surpasses the conventional linear fusion.

[0038] Non-continuous area optimization: first, a Markov random field model is constructed based on the initial depth map, which includes data items and smoothing items. The data items calculate the difference between the current depth estimate and the initial depth estimate using a non-convex function, and the non-convex function is set as the 0.8 power of the absolute value of the depth difference to enhance the robustness to outliers. The smoothing item uses a truncated linear function to calculate the depth difference between adjacent pixels, and when the absolute value of the depth difference is less than a preset threshold, the actual difference value is taken, and when the threshold is exceeded, a fixed truncated value is taken. The overall energy function is expressed by the weighted sum of the data items and the smoothing items, and the smoothing item weight coefficient is controlled by a global parameter; for the detected occlusion boundary area, the smoothing item weight coefficient is set to one tenth of the global weight; finally, the graph cut algorithm is used to minimize the energy function to obtain the optimized depth map.

[0039] Let the initial depth estimate map be , and define the set of adjacent pixel relationships as N, then , : the final depth map to be optimized, where represents the depth value of the pixel point , : the pixel point The initial depth estimate of : a single pixel location identifier in the depth map, : Data term function (non-convex function) acting on the depth estimation deviation , : Smoothing function (truncated linear function) acting on the depth difference of adjacent pixels , : Smoothness weight coefficient, which controls the constraint strength of depth consistency between adjacent pixels. Represents pixels and is the spatial neighborhood relationship, where , , : Smooth term cutoff threshold. If the depth difference exceeds this value, the penalty will no longer increase. The graph cut algorithm is used to solve it, and the smooth weight of the occlusion boundary is specially set. The formula designs the energy function through the Markov random field framework, transforming the depth optimization problem into a mathematical minimization problem, in which the data term forces the optimized depth map to be basically consistent with the initial estimate, using a non-convex function Processing depth deviation, its sublinear characteristics ensure that there is no excessive punishment when there is a large error in the initial estimate, and enhance the robustness to outliers; the smoothing term is achieved by truncating the linear function Distinguish continuous surfaces from real boundaries when the depth difference between adjacent pixels is less than a threshold A linear smoothing constraint is imposed to maintain surface continuity when the depth difference exceeds When the image is smoothed, the jump is allowed to be retained, which physically corresponds to the optical occlusion phenomenon of the real edge of the object; especially for the occluded boundary area, the smoothing weight is reduced to 0.1 times the global weight to avoid the occluded edge being smoothed incorrectly; finally, the overall energy function is minimized by the graph cut algorithm, and the balance between noise suppression and boundary preservation is achieved at the mathematical level. The geometric structure of the depth discontinuous area is accurately reconstructed physically to ensure the correct expression of the object edge sharpness and occlusion relationship.

[0040] Physical constraint iteration: An optimization constraint mechanism is established based on the principle of optical path reversibility. The specific process is as follows: First, the current depth estimate is used to invert the two-dimensional depth map into three-dimensional point cloud data through inverse projection; then the three-dimensional point cloud data is forward projected to each camera perspective coordinate system to generate a synthetic depth map of the corresponding perspective; then the pixel-level difference matrix between the synthetic depth map and the actual observed depth map under all perspectives is calculated; the objective function is constructed as the sum of the squares of the difference matrices for all perspectives; finally, the Levenberg-Marquardt optimization algorithm is used to iteratively update the depth map estimate, and the objective function is minimized through gradient descent adjustment to ensure that the projection consistency of the three-dimensional geometric structure under multiple perspectives is strictly met, focusing on solving the depth inconsistency problem caused by occlusion or noise. Assume that the reference perspective is , any other target perspective is ,but , : Inverse projection operator, the reference view angle Depth map Inversion into three-dimensional point cloud, : Forward projection operator, projects the 3D point cloud to the target perspective The imaging plane, : Target perspective The actual observation depth map, : The square of the Frobenius norm, calculates the sum of the squares of the differences of all elements of the two depth maps. The formula is based on the principle of optical path reversibility: First, the reference view angle Depth map Perform a back-projection operation to invert the two-dimensional depth data into a three-dimensional point cloud; then forward-project the point cloud to the target viewpoint Generate a synthetic depth map; then calculate the synthetic depth map and the actual observed depth map pixel-level differences; ultimately by minimizing all viewing angles The sum of squared differences is used to construct the optimization objective function; its physical meaning is to enforce the cross-view consistency of the three-dimensional geometric structure: if the depth estimation is accurate, then The inverted 3D point cloud is in perspective The projected depth should strictly match the actual observation value; by iterative optimization to minimize the reprojection error, the depth distortion caused by occlusion or noise can be corrected, ensuring that the reconstruction results conform to the optical projection laws and physical interpretability of the real world, thereby improving the geometric accuracy and spatial consistency of the depth map.

[0041] S3 receives the depth feature tensor of S2 and outputs a physically constrained depth map to S4. It improves the quality of object edge reconstruction through multi-scale confidence fusion to adapt to changes in viewing distance. Its physical consistency iteration ensures that the depth map conforms to the laws of optical propagation.

[0042] S4: Dynamic wavefront modulation display control: composed of phase distribution calculation, aberration real-time compensation and multi-plane fusion, based on depth field for wavefront phase calculation and Helmholtz decomposition, real-time monitoring and compensation of optical aberration through embedded wavefront sensor, finally realizing multi-focal plane light field fusion by using angular spectrum diffraction theory.

[0043] Phase distribution calculation: first receive depth map data from hierarchical depth optimization engine output, then calculate the initial value of wavefront phase of each pixel point based on physical optical principle, this process converts depth value into phase value according to the phase delay characteristics of light wave propagation in medium; let the refractive index of medium be , then the wavefront phase value , : the wavefront phase value at the position coordinate , describes the phase offset of light wave at this point; this formula is based on the phase delay principle of light wave propagation in medium, the core logic is to convert depth information into light wave phase modulation amount, when light wave propagates in medium, it produces a complete phase period change of radian per wavelength distance, and the refractive index of medium changes the equivalent path length of light wave propagation, so the formula determines the phase change rate per unit depth through the ratio of light wave wavelength parameter and medium refractive index parameter, and then multiplies the distance value of depth map to establish a linear mapping relationship from depth information to wavefront phase, its physical meaning is to reconstruct the light field characteristics of the original three-dimensional scene, the spatial light modulator applies delay to the incident light wave according to this phase value, so that the light rays re-converge in space to form stereoscopic image points, different depth positions correspond to different phase modulation amounts, finally realizing the correct intersection of light rays at each viewing angle in three-dimensional space to produce real stereoscopic visual effect when observed with naked eyes; Then calculate the two-dimensional gradient field of the phase field, respectively solve the phase partial derivatives in X and Y directions to construct the gradient vector field; then reconstruct the gradient field into continuous phase distribution by solving Poisson equation, use multi-grid iteration algorithm to maintain convergence stability at the occlusion boundary; then implement Helmholtz decomposition on the reconstructed phase to separate the divergence component and the irrotational component, only keep the irrotational gradient component as the final modulation reference, finally output the continuous phase distribution graph conforming to the wavefront propagation law to the aberration real-time compensation process; this phase graph directly determines the generation of control signal of subsequent spatial light modulator, the whole process strictly follows the principle of optical reversibility to ensure the accurate reconstruction of light field characteristics of three-dimensional scene at the display end; let the Helmholtz decomposition of phase field be , : original phase field (vector field), represents the phase distribution of each point in space, : irrotational component (gradient of scalar potential function), where is the scalar potential function, describes the conservative field component of phase, : passive component (curl of vector potential function), where is the vector potential function, describing the non-conservative field component of phase; this formula is based on the principle of vector field decomposition: any continuous vector field can be uniquely decomposed into a non-rotating component (conservative field) and a non-divergent component (curl field), which physically separates the physically achievable part from the non-physical distortion component in the phase field, where the non-rotating component corresponds to the conservative field characteristics required for actual light wave propagation, and its curl is always zero, ensuring that the light propagation path satisfies the reversibility principle; the curl component reflects the discontinuity of the occlusion boundary left by the depth map optimization, which is manifested as a non-physical vortex structure that will cause ghosting and crosstalk distortion in light field reconstruction. Through this decomposition, the system can accurately extract the pure gradient phase field that meets the wave equation constraints, while filtering out the noise components caused by depth discontinuity, reducing the wavefront error to the minimum limit of physical realizability. This process is achieved through the divergence-curl iteration method.

[0044] Aberration real-time compensation: first, deploy a micro wavefront sensor at each of the four edge positions of the display screen to continuously monitor the phase distribution state of the output light field, and the sensor synchronously collects the actual wavefront distortion data at a sampling frequency of 120Hz; then input the collected distortion data into the pre-set aberration mathematical model, which uses a combination of fifteen standard orthogonal basis functions to describe common aberration types; subsequently, the least squares optimization algorithm is used to dynamically solve the corresponding coefficient combination of the current aberration state, and a sliding window filtering mechanism is introduced to eliminate the influence of instantaneous measurement noise; then the aberration coefficients obtained by solving are converted into corresponding phase compensation amounts, and are superimposed on the original phase distribution map to generate the corrected modulation signal; finally, the real-time updated modulation signal is transmitted to the driving unit of the spatial light modulator through a high-speed data interface, forming a closed-loop control system to continuously suppress the aberration changes caused by environmental disturbances, optical device thermal drift, etc., ensuring that the output light field always meets the theoretical reconstruction requirements, and the corrected phase data is directly output to the multi-plane fusion process for subsequent processing; let the normalized radial coordinate be , the range , representing the relative distance from the center to the edge of the display, and the azimuthal coordinate be , the range , representing the angular position on the display plane, then , : compensated phase distribution function, representing the phase correction amount required for each coordinate point on the display plane, : weight coefficient of the kth Zernike mode, determining the compensation strength of the aberration mode, : the kth Zernike orthonormal basis function, describing a specific type of wavefront aberration mode; the formula is based on the principle of Zernike polynomial orthogonal decomposition, first the phase aberration data of the actual light field is collected in real time by the wavefront sensor, then the aberration field is mapped into the mathematical space spanned by 15 standard orthonormal basis functions, each basis function strictly corresponds to a specific physical aberration mode; then the least squares optimization algorithm is used to dynamically solve the weight coefficients of each basis function, and the process converts the discrete sampling point data of the sensor into a continuous compensation function through matrix operation; the final generated compensation phase function is essentially a digital twin modeling of the system inherent aberration, its physical meaning is to actively construct a compensation wave surface with the same amplitude and opposite phase as the aberration wavefront, when this compensation quantity is loaded into the optical path through the spatial light modulator, it can accurately offset the composite aberrations such as defocus, astigmatism, coma and other factors caused by optical element processing error, mechanical assembly stress, environmental temperature drift, etc., at the same time, by updating the weight coefficients in real time, the system is given the ability to dynamically adapt to environmental disturbances, ensuring that the wavefront accuracy of the output light field always meets the visual diffraction limit requirements.

[0045] Multi-plane fusion composition: first, the axial positions of three virtual focal planes are dynamically determined according to the scene depth distribution range, and the three planes correspond to the typical focusing distances of the foreground, middle ground and background respectively; then the diffraction propagation calculation is independently performed for each virtual focal plane, and the intermediate light field distribution corresponding to the depth is generated through the preset light wave transmission model; then the adaptive fusion weight is calculated based on the statistical characteristics analysis of the depth map, and the weight is dynamically adjusted by the entropy value of the depth distribution to ensure the priority presentation of the detailed rich area; finally, the intermediate light fields of the three focal planes are weighted and superimposed, and a phase alignment correction mechanism is introduced in the superposition process to eliminate interference artifacts between different planes, generating the final full-depth light field output consistent with the human visual characteristics; the entire fusion process is strictly synchronized with the display refresh rate, and its output directly drives the spatial light modulator to form a naked-eye three-dimensional visual effect without auxiliary equipment, and through real-time monitoring of scene depth changes, the plane position and fusion parameters are dynamically updated to ensure visual continuity in motion scenes. Set the three virtual focal planes as , the light wave transmission model , : the three-dimensional light field intensity distribution of the final fusion output, : the adaptive fusion weight of the kth plane, : the diffraction propagation operator, : initial complex amplitude distribution of the spatial light modulator plane; the formula design is first based on the angular spectrum diffraction theory to accurately simulate the physical process of light wave propagation from the spatial light modulator to each virtual focal plane, and the core logic is to calculate the light field distribution at different depths through the complex amplitude transfer function; the weight calculation term introduces the depth distribution entropy as a dynamic adjustment parameter, and the physical meaning is to quantify the spatial complexity of the scene, and automatically increase the fusion ratio of the corresponding focal plane to enhance the detail restoration when the depth changes sharply; the fusion process adopts the additive superposition principle, and the physical nature is to follow the linear superposition characteristics of light wave interference, and the complex amplitude fields of the three planes are synthesized according to the entropy weight coefficient; the specially designed phase alignment correction term eliminates the multi-plane interference fringes by compensating the propagation path difference, and the physical meaning is to maintain the wavefront continuity to avoid visual artifacts; the whole formula system dynamically optimizes the number and position of focal planes through the depth self-adaptive mechanism, and the fundamental physical goal is to expand the depth of field while maintaining the diffraction limit resolution, and finally realize the full-field light field reconstruction conforming to the Helmholtz wave equation.

[0046] The S4 receives the light field modulation signal generated by the optimized depth map of the S3, directly determines the light field reconstruction accuracy of the naked eye 3D through wavefront phase decomposition and multi-plane fusion, and the real-time aberration compensation mechanism enhances the three-dimensional display stability of the system in complex environment.

[0047] The application starts from collecting the environment of the controlled scattering medium light field, which is filled with silica suspended particles of specific density and particle size as a controllable aerosol medium in a closed darkroom through environment parameter configuration, uses an LED array to generate a structured light coding sequence and projects it through a projection lens in combination with an active light source coding, simultaneously implements multi-view synchronous acquisition to realize hardware-level synchronous triggering and data acquisition using a camera array, and outputs a multi-view light field sequence data carrying scattering characteristics; then enters the light field data analysis stage of scattering feature enhancement, first performs medium scattering modeling based on Mie scattering theory to construct a light transmission physical model, then processes image data through frequency domain-space domain joint decoupling, including windowed Fourier transform to separate frequency domain components and non-local mean filtering to optimize space domain features, finally uses a nonlinear depth feature extraction to extract cross-view geometric correlation features using a cascade neural network architecture, and outputs a depth feature tensor; subsequently, a hierarchical depth optimization engine performs multi-scale confidence evaluation to construct a multi-scale depth map gradient confidence and performs weighted fusion, implements non-continuous region optimization to process depth discontinuous regions based on a Markov random field model and solves it using a graph cut algorithm, and finally performs physical constraint iteration to calculate a reprojection error using the reversibility principle of light path and updates the depth map through an optimization algorithm, and outputs a physically constrained depth map; finally, dynamic wavefront modulation display control calculates the phase distribution based on the depth map to convert the depth value into a wavefront phase and implements Helmholtz decomposition, then uses a wavefront sensor to monitor distortion through real-time aberration compensation and applies a Zernike model to generate a compensation phase, finally performs multi-plane fusion to determine the virtual focal plane position, calculate the diffraction propagation and weighted superposition of the light field, generates a light field modulation signal to drive a spatial light modulator to realize naked-eye 3D display, and the entire process ensures the accurate execution of light field data collection, depth feature extraction optimization and display control.

[0048] Secondly: the drawings of the disclosed embodiments only involve the structures involved in the disclosed embodiments, other structures can refer to the general design, and the same embodiments and different embodiments of the application can be combined with each other under the condition of no conflict; Finally: the above only describes the preferred embodiments of the application and is not used to limit the application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application should be included in the protection scope of the application.

Claims

1. A naked-eye 3D display system based on a light field reconstruction algorithm, characterized in that: include: S1: Controlled scattering medium light field acquisition environment: This consists of environmental parameter configuration, active light source coding, and multi-view synchronous acquisition. By controlling the physical properties of the aerosol medium, combining a structured light coding scheme, and using a multi-camera array for time synchronization, a special light field data acquisition environment is constructed. S2: Light field data analysis with enhanced scattering features: This includes medium scattering modeling, frequency-space domain joint decoupling, and nonlinear deep feature extraction. A light transmission physical model is established based on Mie scattering theory. Frequency domain signal separation and reconstruction are achieved through windowed Fourier transforms. Spatial domain features are processed using non-local mean filtering. Finally, a cascaded neural network architecture is used to extract cross-viewpoint geometric correlation features. S3: Hierarchical depth optimization engine: This engine performs multi-scale confidence assessment, discontinuous region optimization, and physical constraint iteration. It first constructs a Gaussian pyramid for multi-scale gradient confidence fusion, then uses an improved Markov random field model to process depth discontinuities, and finally introduces optical path reversibility constraints for iterative optimization. S4: Dynamic wavefront modulation display control: It consists of phase distribution calculation, real-time aberration compensation and multi-plane fusion. It performs wavefront phase calculation and Helmholtz decomposition based on the depth field, monitors and compensates optical aberrations in real time through embedded wavefront sensors, and finally uses angular spectrum diffraction theory to realize multi-focal plane light field fusion.

2. The naked-eye 3D display system based on the light field reconstruction algorithm according to claim 1, characterized in that: The environmental parameter configuration includes: A closed darkroom is filled with suspended silica particles with a particle size ranging from 1 μm to 5 μm and a density controlled at 0.5 g / m³ to 1.5 g / m³ as a controllable aerosol medium. The refractive index of the medium is set to 1.45 and a floating tolerance of ±0.05 is allowed.

3. The naked-eye 3D display system based on the light field reconstruction algorithm according to claim 1, characterized in that: The active light source coding includes: The light source utilizes an LED array consisting of 450nm blue, 520nm green, and 630nm red LEDs arranged in a 1:1:1 staggered ratio. A programmable current driver synchronously switches independent coding sequences of the three wavelengths at a frame rate of 120fps. Each coding cycle consists of a 12-bit binary Gray code sequence. The projection lens assembly is equipped with an f=8mm fisheye lens to achieve a 120-degree field of view. The coding pattern generates spatially unique two-dimensional structured light stripes based on the De Bruijn sequence, with a minimum line width of 0.5mm. Pattern switching is synchronized with the camera and exposure. The FPGA trigger circuit has a timing jitter of less than 1μs. The LED drive current peaks at 5A and maintains an 80% duty cycle in pulse-width modulation mode. The coded data is preloaded into the light source controller cache via a Gigabit Ethernet interface. During implementation, the LED junction temperature is monitored in real time, and the cooling fan speed is dynamically adjusted to maintain a temperature range of 45±2°C. A closed-loop light intensity feedback mechanism adjusts the drive current based on the photodiode sampling value, ensuring that the projection brightness fluctuation does not exceed 5% of the nominal value.

4. The naked-eye 3D display system based on the light field reconstruction algorithm according to claim 1, characterized in that: The multi-view synchronous acquisition includes: A 5×5 rectangular array is used to arrange 25 global shutter CMOS industrial cameras. The resolution of each camera is set to 4096×2160 pixels and is equipped with a 16mm focal length fixed focus lens. The baseline spacing of the camera array is controlled to 60mm. All cameras are hardware-level synchronized and triggered by the FPGA main control board. The trigger signal is directly connected to the GPIO trigger port of each camera using a star topology. The timing jitter is controlled within the range of ±0.2μs. The exposure time is uniformly set to 1 / 240 second and maintains a 2:1 integer multiple relationship with the 120fps refresh rate of the active light source encoding. The image data is transmitted via Camera Link. The data is transmitted to the pre-processing server via the HS interface. Each transmission link uses single-mode optical fiber to ensure that the total bandwidth of the 25 video streams is no less than 375Gbps. During the acquisition process, a non-uniformity correction algorithm is applied in real time to eliminate the dark current differences of each sensor. At the same time, an integrated temperature sensor is used to monitor the operating temperature of each camera and dynamically adjust the power of the thermoelectric cooler to maintain the sensor temperature within the range of 30±0.5℃. Before each acquisition, a high-precision ceramic calibration plate is used for sub-pixel optical distortion correction. The error of the spacing between the feature points of the calibration plate is less than 1μm. The final raw data is stored in 16-bit RAW format as a multi-frame time series container file, and the timestamp and spatial pose matrix data accurate to the nanosecond level are embedded in the file header.

5. The naked-eye 3D display system based on the light field reconstruction algorithm according to claim 1, characterized in that: The medium scattering modeling includes: First, a light intensity distribution calculation model is constructed based on Mie scattering theory. This process obtains three core input parameters: the raw light intensity data of the active light source in the multi-view acquisition system, the pre-calibrated physical properties of the aerosol medium, including the particle size distribution range of the silica suspended particles and the medium density, and the scattering angle observations recorded by 25 synchronized cameras. During the model calculation stage, the Mie scattering principle is applied to establish a mapping relationship between the incident light intensity and the scattered light intensity. First, the Mie coefficient is calculated based on the real part of the complex refractive index of the suspended particles, 1.45, and the imaginary part, 0.

05. The size parameters of the scattering particles and the wavelength of the incident light are then used to determine the upper limit of the order of the series expansion. Then, a recursive algorithm is used to generate a numerical solution to the angular distribution function, and finally, a theoretical scattered light intensity distribution map is output for each camera perspective.

6. The naked-eye 3D display system based on the light field reconstruction algorithm according to claim 1, characterized in that: The frequency domain-spatial domain joint decoupling includes: First, a windowed Fourier transform is performed on the raw image captured from each camera viewpoint. An adjustable local analysis window is used to calculate frequency components region by region in the spatial domain, decomposing the image into components of different spatial frequencies. Next, a feature separation operation is performed in the frequency domain. Based on the scattering physics model, low-frequency components representing direct illumination and high-frequency components carrying scattering characteristics are identified and separated. A dynamic energy threshold is set to retain high-frequency components that represent 5% to 15% of the total energy as depth information carriers, while the remaining frequency components are suppressed. The filtered high-frequency components are then converted back to spatial domain representation and subjected to an improved non-local mean filtering process. This filtering process calculates the similarity weights between image blocks and performs adaptive smoothing operations on the selected high-frequency components while retaining edge features.

7. The naked-eye 3D display system based on the light field reconstruction algorithm according to claim 1, characterized in that: The nonlinear depth feature extraction includes: First, a three-level cascade deep learning architecture is constructed. The first level uses a three-layer convolutional neural network to process the preprocessed multi-view image blocks. The first layer uses a 7×7 convolution kernel to extract large-scale local geometric features to generate a 128-channel feature map. The second layer uses a 5×5 convolution kernel to compress the features and reduce the dimensionality to 64 channels. The third layer uses a 3×3 convolution kernel to capture texture details and output a 32-channel feature tensor. All convolution layers are followed by LeakyReLU activation function and batch normalization operation; then the second level introduces a graph convolution network to establish the correlation between perspectives, takes 25 camera perspectives as graph nodes and constructs an adjacency matrix according to the camera pose parameters, and defines the nodes specifically. The edge weight is the inverse of the Euclidean distance between the camera projection centers multiplied by the cosine of the optical axis angle. The features of neighboring nodes are iteratively aggregated through three layers of graph convolution to achieve cross-view information fusion. Finally, the third level uses a temporal long short-term memory network to process the dynamic coded light sequence. The 12 consecutive frames of Gray code images of each perspective are used as the time series input. The memory gating mechanism designs a phase shift sensitive unit, which calculates the local area cross-correlation value of adjacent frames as the gating signal input. The network output is connected to the full convolution decoding layer to map the spatiotemporal features into the initial depth map. The final output layer uses a 1×1 convolution kernel with a Sigmoid activation function to generate a normalized depth prediction value.

8. The naked-eye 3D display system based on the light field reconstruction algorithm according to claim 1, characterized in that: The multi-scale confidence assessment includes: First, three depth map representations of different scales are constructed on the original depth map, and multi-scale expression is achieved through Gaussian blur and downsampling operations; then the spatial gradient amplitude is calculated independently on each scale depth map, and the gradient amplitude is converted into a local confidence measure. The specific method is to perform negative exponential mapping on the square of the gradient amplitude and take the complement, where the scale coefficient of the exponential function is a preset constant control; Finally, the confidence metrics at the three scales are geometrically weighted fused. The weight coefficient of each scale is pre-set according to the actual application requirements. The fusion result is used as the final confidence evaluation value of each pixel position for subsequent optimization decisions.

9. The naked-eye 3D display system based on the light field reconstruction algorithm according to claim 1, characterized in that: The discontinuous region optimization includes: Firstly, a Markov random field model is constructed based on the initial depth map. The model consists of two parts: data term and smoothing term. The data term uses a non-convex function to calculate the difference between the current depth estimate and the initial depth estimate. The non-convex function is set to the 0.8 power of the absolute value of the depth difference to enhance the robustness to outliers. The smoothing term uses a truncated linear function to calculate the depth difference between adjacent pixels. When the absolute value of the depth difference is less than the preset threshold, the actual difference value is taken. When it exceeds the threshold, a fixed truncation value is taken. The overall energy function is expressed by the weighted sum of the data term and the smoothing term, where the smoothing term weight coefficient is controlled by a global parameter. For the detected occlusion boundary area, its smoothing term weight coefficient is set separately to one tenth of the global weight. Finally, the graph cut algorithm is used to minimize the energy function to obtain the optimized depth map.

10. The naked-eye 3D display system based on the light field reconstruction algorithm according to claim 1, characterized in that: The physical constraint iteration includes: An optimization constraint mechanism is established based on the principle of optical path reversibility. The specific process is as follows: first, the current depth estimate is used to invert the two-dimensional depth map into three-dimensional point cloud data through a back projection operation; then the three-dimensional point cloud data is forward projected to the coordinate system of each camera view to generate a synthetic depth map of the corresponding view; then the pixel-level difference matrix between the synthetic depth map and the actual observed depth map under all view angles is calculated; the objective function is constructed as the sum of the squares of the difference matrices of all view angles; finally, the Levenberg-Marquardt optimization algorithm is used to iteratively update the depth map estimate, and the objective function is minimized through gradient descent adjustment.

Images