A multi-domain voxel reconstruction three-dimensional display method and device

Abstract

The present application relates to the field of computational imaging and three-dimensional display, specifically to a multi-domain voxel reconstruction three-dimensional display method and device, comprising the following steps: S1: voxel layer feature extraction and convolution reconstruction: using convolutional neural network to extract features and repair structure of input voxel layer; S2: voxel layer fusion based on learning weight: through learning weight distribution and edge compensation mechanism, realizing natural interlayer transition and structure enhancement; S3: in focus direction, perception fusion weight modeling: according to observation focus depth and depth difference of each voxel layer, automatically distributing fusion weight, making focusing layer clear and defocusing layer smooth attenuation, thereby strengthening spatial stereoscopic effect; S4: multi-domain fusion and adaptive compensation output: in the fusion stage, introducing depth residual term and spatial consistency constraint, realizing cross-layer smooth transition and overall brightness self-balancing.

Description

Technical Field

[0001] This invention relates to the field of computational imaging and 3D display technology, specifically to a multi-domain voxel reconstruction 3D display method and apparatus. Background Technology

[0002] With the rapid development of 3D display technology, 3D display methods based on voxel reconstruction and focus awareness have gradually become a research hotspot due to their advantages in realistic viewpoint, depth perception, and image quality. However, existing 3D display technologies still have shortcomings in multi-domain voxel reconstruction, dynamic focus adjustment, and complex scene processing, which affect display effects and user experience.

[0003] Existing technology discloses a "depth scanning naked-eye 3D display method" (CN111240036B). This patent achieves true 3D object display with a realistic viewing angle and self-occlusion effect by moving a light-emitting screen back and forth along the main optical axis, combined with the design of concave mirrors and imaging mirrors. However, this technical solution mainly relies on mechanical devices to control the movement of the light-emitting screen, resulting in a slow system response speed and difficulty in handling complex dynamic scenes in real time. In addition, this method does not involve intelligent optimization algorithms based on convolutional neural networks, and cannot achieve efficient reconstruction and rendering of multi-domain voxel data, limiting its application in high-precision 3D display. Another existing technology discloses a "flow field multi-section schlieren synchronous display system and method" (CN114923659B). This patent combines focused schlieren technology with light field imaging technology to acquire schlieren images of any focused section in the flow field at the same time, and completes 3D reconstruction through image overlay. However, this technical solution is mainly aimed at the field of optical flow display, lacks the ability to process multi-domain voxel data in depth, and does not introduce a focus perception mechanism, resulting in insufficient accuracy in expressing depth information in complex 3D scenes. In addition, this method is highly dependent on hardware devices, has high system costs, and is difficult to meet the needs of large-scale promotion.

[0004] The above problems indicate that existing 3D display technologies still have certain shortcomings in terms of dynamic response speed, multi-domain voxel reconstruction capability, focus perception optimization, and system cost control. Summary of the Invention

[0005] This invention discloses a multi-domain voxel reconstruction three-dimensional display method and apparatus, which aims to improve the above-mentioned technical problems.

[0006] The present invention adopts the following scheme: a multi-domain voxel reconstruction 3D display method, comprising the following steps: S1: voxel layer feature extraction and convolutional reconstruction: using a convolutional neural network to extract features and repair the structure of the input voxel layer; S2: voxel layer fusion based on learned weights: through learned weight allocation and edge compensation mechanism, natural inter-layer transition and structural enhancement are achieved; S3: perceptual fusion weight modeling in the focal direction: according to the difference between the depth of the observed focal point and the depth of each voxel layer, the fusion weight is automatically allocated to make the focused layer clear and the defocused layer smoothly attenuated, thereby enhancing the spatial stereoscopic effect; S4: image output after multi-domain fusion and adaptive compensation: in the fusion stage, depth residual terms and spatial consistency constraints are introduced to achieve smooth cross-layer transition and overall brightness self-balancing.

[0007] Preferably, step S1 includes the following steps: (1) Input voxel sequence: Input a multi-plane image voxel sequence I(x, y, l), where l ∈ [1, 32] represents the depth layer index, each layer corresponds to a different spatial depth slice in the scene, and x, y represent the horizontal and vertical coordinates of the planar image; (2) Multi-branch convolution feature extraction: Each branch corresponds to a convolution kernel of different scales. (2) Used to extract local texture and global geometric features, and establish cross-depth dependencies between layers; (3) In the feature fusion stage, a two-dimensional Laplacian operator is introduced. The formula used to enhance the feature gradient is: ,in Indicates the first Layer input voxel feature map, This represents the weighted convolution operation of the i-th scale convolution branch on the features of adjacent layers. This represents the second-order spatial Laplacian enhancement of the features of this layer. The structural reinforcement factor is... (4) Perform feature backpropagation and cross-layer reconstruction: Introduce residual units to realize cross-layer feature feedback and upsampling repair: ,in This is the output feature for reconstructing the l-th layer. This indicates an upsampling operation, which modifies the features from the previous layer... Interpolate to the current resolution. This represents the cross-layer fusion balance factor, used to adjust the contribution of inter-layer residuals. This represents the feature residual between the current layer and the previous layer, used to capture the subtle differences caused by depth variations.

[0008] Preferably, the establishment of cross-depth dependency in step (2) of S1 includes the following steps: by feeding adjacent depth layers together into the attention matrix, the features of each layer not only depend on itself, but also on other depth layers, specifically as follows: In the deep attention scheme, the input voxel I(x,y,l) of each depth layer first passes through the convolution kernel W i Extract local geometric features to obtain the corresponding query vector Q. l Key vector Kl and value vector V l Then the attention matrix A(l,k) is based on Q l and K k The similarity learning determines the degree of dependency between different depth layers, and this is used to apply the value vector V of different depth layers. k Weighted fusion is performed to establish long-range cross-depth feature dependencies, where the attention matrix expression is as follows:

[0009]

[0010] Where l and k are depth layer indices, and k is the other depth layer that is associated with the l-th layer for calculation;

[0011] The final generated layer l fusion feature for:

[0012]

[0013] Convolution kernels generate Q features Used to measure the depth structure of this layer:

[0014]

[0015] Convolution kernel generates K features Used to characterize the geometric features that can be provided:

[0016]

[0017] The convolutional kernel generates V features, which are used for the feature output after cross-depth fusion.

[0018]

[0019] convolution kernel All are weighted parameters. Feature encoding from convolutional networks The transpose of the query vector is used to calculate cross-depth feature similarity, and Softmax() normalizes the similarity across all depth layers into attention weights.

[0020] Preferably, step S2 includes the following steps:

[0021] (1) Calculate the fusion weights for each layer : Targeting pixels At the given location, the learned weight network adaptively estimates the fusion weights for layers l at different depths. This is to reflect the credibility and salience of each layer in the local area;

[0022] (2) Weighting the multi-layer voxel results by weight: Based on the voxel reconstruction results of each layer... With weight Perform a weighted average operation to generate a fused image;

[0023] (3) Edge-guided blur region restoration: In the fused image, edge-guided items are used... Enhancement and repair of blurred areas are performed by strengthening edge energy response through inverse frequency domain transform, restoring structural clarity and contour sharpness. The core expression is:

[0024]

[0025] in To produce the merged output image, This is the result of voxel reconstruction of layer l. These are the fusion weights, reflecting the contribution of this layer at pixel p. It is the inverse Fourier transform, used to recover edge energies in the frequency domain. The gradient distribution of the edge map, is the edge enhancement coefficient, which controls the repair intensity, and L is the total number of layers.

[0026] Preferably, step S3 includes the following steps:

[0027] (1) Set focus and depth parameters: Let the current focus depth be . The center depth of each voxel layer is The system determines the sharpness contribution of each layer by calculating the depth difference between different layers and the focus.

[0028] (2) Calculate the focus response weights Based on the focusing behavior of the human eye, a Gaussian focus response function is constructed, and a depth gradient term is introduced to reflect the focal length sensitivity of the scene edge;

[0029] (3) Perform voxel focusing fusion: The results of multiple voxels are weighted and fused according to the focus response weights, so that the features at the focus point are enhanced and the defocus points are naturally attenuated, so as to obtain a focusing effect consistent with human eye perception. The focus weighting function is as follows:

[0030]

[0031] in Indicates the first Layer focus response weights, The normalization constant is For the first Layer center depth, For the current focus depth, This represents the depth of field width parameter. The depth gradient influence coefficient. The depth gradient magnitude represents the local depth change rate.

[0032] The focus-weighted fusion result is:

[0033]

[0034] in The result of focus-weighted fusion. Indicates the first Layer-weighted fusion image, To indicate the first Layer focus response weights, This represents the total number of voxel layers.

[0035] Preferably, step S4 includes the following steps: frequency domain energy enhancement: [The text abruptly ends here, so the translation stops.] Convert to frequency domain:

[0036]

[0037] Inverse transformation

[0038]

[0039] Energy-enhanced operations suppress low-frequency blurring and amplify high-frequency detail components to restore structural texture and edge sharpness. This represents the merged image after focus weighting. express The two-dimensional discrete Fourier transform, For pixel coordinates, u and v are frequency domain indices or normalized frequencies;

[0040] (1) Depth residual compensation: Based on frequency domain optimization, a depth residual term is introduced to correct the geometric deviation between voxel layers, thereby improving depth continuity and spatial alignment accuracy.

[0041] (2) Gradient consistency optimization: Apply gradient consistency constraints in the spatial domain to adaptively adjust brightness and structure, ensuring smooth edge transition and maintaining global illumination balance;

[0042] (4) Image output:

[0043] in: Indicates Fourier transform, For the depth Laplace term, This is the structural compensation coefficient. This indicates the final output image. These represent the Fourier transform and the inverse transform, respectively. The Laplacian term of the depth map is used to enhance the edges of structures.

[0044] A multi-domain voxel reconstruction 3D display device is provided, which uses the 3D display method described above for display.

[0045] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0046] This invention provides a multi-domain voxel reconstruction 3D display method based on convolutional neural networks and focus perception. It aims to optimize the voxel reconstruction process through intelligent algorithms, improve the accuracy of depth information representation by combining focus perception mechanism, and reduce system complexity and cost, thereby meeting the needs of modern 3D display technology for efficiency, accuracy and economy. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0048] Figure 1 This is a schematic diagram illustrating the working principle of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. 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. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. 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.

[0050] Example

[0051] Combination Figure 1This invention provides a multi-domain voxel reconstruction three-dimensional display method, comprising the following steps:

[0052] Step 1: Voxel layer feature extraction and convolutional reconstruction

[0053] Since multi-plane imaging (MPI) uses discrete depth planes to represent scenes, it is essentially a 2.5D representation. Discrete depth sampling can easily lead to problems such as discontinuous object edges, blurred textures, and occlusion errors. To address this, this invention first utilizes a convolutional neural network (CNN) to extract features and repair the structure of the input voxel layer, thereby improving depth continuity and texture fidelity.

[0054] (1) Input voxel sequence

[0055] Input a multiplane image (MPI) voxel sequence I(x, y, l), where l ∈ [1, 32] represents the depth layer index, and each layer corresponds to a different spatial depth slice in the scene.

[0056] (2) Multi-branch convolution feature extraction

[0057] To fully capture the spatial correlation between different layers, a multi-branch convolutional structure is used to extract features from the voxel maps of each layer. Each branch corresponds to a convolutional kernel of a different scale. This is used to extract local texture (small kernel) and global geometry (large kernel) features, and to establish cross-depth dependencies between layers. Establishing cross-depth dependencies involves the following steps: by feeding adjacent depth layers together into the attention matrix, the features of each layer depend not only on itself but also on other depth layers, specifically as follows: In the deep attention scheme, the input voxel I(x,y,l) of each depth layer first passes through the convolution kernel W. i Extract local geometric features to obtain the corresponding query vector Q. l Key vector K l AND value vector V l Then the attention matrix A(l,k) is based on Q l and K k The similarity learning determines the degree of dependency between different depth layers, and this is used to apply the value vector V of different depth layers. k Weighted fusion is performed to establish long-range cross-depth feature dependencies, where the attention matrix expression is as follows:

[0058]

[0059] Where l and k are depth layer indices, and k is the other depth layer that is associated with the l-th layer for calculation;

[0060] The final generated layer l fusion feature for:

[0061]

[0062] Convolution kernels generate Q features Used to measure the depth structure of this layer:

[0063]

[0064] Convolution kernel generates K features Used to characterize the geometric features that can be provided:

[0065]

[0066] The convolutional kernel generates V features, which are used for the feature output after cross-depth fusion.

[0067]

[0068] convolution kernel These are all weight parameters, which, after expansion, are multiplied with the input voxel layer matrix, and the result is expressed in matrix form. , , This corresponds to the Query (used to "initiate a query," determining what structure this layer is looking for), the Key (used to "provide information," used for matching with the Query), and the Value feature (the actual feature ultimately used for weighted fusion). Then, cross-depth similarity is calculated through matrix inner product and Softmax normalization is applied to obtain the depth attention weights. Finally, matrix multiplication completes the cross-depth feature fusion of layer l. 。

[0069] in Feature encoding from convolutional networks The transpose of the query vector is used to calculate cross-depth feature similarity, and Softmax() normalizes the similarity across all depth layers into attention weights.

[0070] (3) Cross-layer feature fusion and structural enhancement

[0071] In the feature fusion stage, a two-dimensional Laplacian operator is introduced. The feature gradient is enhanced to highlight object edges and regions of abrupt depth changes, thereby strengthening spatial structural information.

[0072] Core formula:

[0073]

[0074] (The notation k is the radius of the neighboring depth layer, i.e., a positive integer, indicating that the range of neighboring depth layer indices participating in the weighted convolution when generating the output at the l-th layer is from l.) k to l+k. (To avoid index out-of-bounds, when l+i exceeds [1,L] (where L is the total number of voxel layers), it is handled in one of the following ways: the index that exceeds the range is truncated (for example, the item l+i<1 is regarded as l+i=1, and the item l+i>L is regarded as l+i=L)).

[0075] in: This represents the input voxel feature map of the l-th layer. ai represents the weighted convolution operation of the i-th scale convolution branch on the features of adjacent layers. This represents a second-order spatial Laplacian enhancement of the features of this layer, used to strengthen the gradient at edges and depth changes. This is the structural reinforcement coefficient, used to control the degree of edge reinforcement. (.) represents a non-linear activation function (such as ReLU) to ensure that the network has non-linear mapping capabilities.

[0076] This formula essentially achieves cross-layer feature aggregation and edge-enhanced convolution: the system extracts spatial context features by sliding multi-scale convolution kernels between neighboring layers and introduces the Laplacian operator to enhance structural texture, enabling the model to simultaneously possess the feature representation capabilities of "local refinement" and "global continuity".

[0077] (4) Feature backhaul and cross-layer reconstruction

[0078] To further enhance interlayer continuity, residual units are introduced to achieve cross-layer feature feedback and upsampling repair.

[0079]

[0080] in: The output features are reconstructed for the l-th layer. This indicates an upsampling operation, which modifies the features from the previous layer. Interpolate to the current resolution. This represents the cross-layer fusion balance factor, used to adjust the contribution of interlayer residuals. This represents the feature residual between the current layer and the previous layer, used to capture the subtle differences caused by depth variations.

[0081] This formula achieves self-correction of inter-layer features, enabling the depth boundary and texture details to be continuously connected in the vertical direction, avoiding the "layer jump" and "overlapping artifact" problems caused by traditional MPI discrete slicing.

[0082] Technical Results: Through the above steps, the system achieves continuous depth reconstruction and texture detail restoration of discrete voxel layers. This module not only enhances the geometric accuracy of depth boundaries but also provides a high-fidelity voxel input foundation for subsequent focus weighting and multi-domain compensation.

[0083] Step 2: Voxel layer fusion based on learned weights

[0084] Content: Features in different voxel layers exhibit spatial misalignment and brightness inconsistencies. Through learned weight allocation and edge compensation mechanisms, natural interlayer transitions and structural enhancements can be achieved.

[0085] (1) Calculate the fusion weights for each layer

[0086] For the location of pixel p, the learned weight network adaptively estimates the fusion weights of layers l at different depths. This is to reflect the credibility and salience of each layer in the local area.

[0087] (2) Weight the multi-voxel results according to the weights.

[0088] Based on the voxel reconstruction results of each layer With weight A weighted average operation is performed to form the fused image. This process ensures that the high-confidence layer is larger, thereby suppressing inter-layer misalignment and occlusion interference.

[0089] (3) Edge-guided blur area repair

[0090] In the fused image, edge guiding terms are used. The blurred areas are enhanced and repaired by strengthening the edge energy response through inverse frequency domain transformation, thereby restoring structural clarity and contour sharpness.

[0091] Core expression:

[0092]

[0093] in: This is the output image after fusion. It is the first Voxel reconstruction results of the layer. These are fusion weights, reflecting the layer's performance on pixels. The contribution of the department. It is the inverse Fourier transform, used to recover edge energy in the frequency domain. This represents the gradient distribution of the edge map. The edge enhancement coefficient controls the repair intensity. This refers to the total number of layers (e.g., a 32-layer voxel slice).

[0094] The first part of the expression ensures geometric continuity between layers, while the second part enhances the edges in the frequency domain, resulting in a fusion result that is both smooth and clear. It maintains spatial brightness stability while suppressing blurring and tortuosity caused by interlayer misalignment.

[0095] Technical effects: This step achieves the following improvements by introducing a triple mechanism of learning weight fusion, frequency domain edge repair, and gradient consistency constraints: significantly reducing interlayer breaks and illumination discontinuities; ensuring geometric consistency and sharp edges in the fusion result; and enhancing the sense of spatial detail and visual balance.

[0096] Step 3: Focus Direction Awareness Fusion Weight Modeling

[0097] Content: To enhance the focus effect from the human eye's primary viewpoint and simulate realistic focusing behavior, this invention introduces a focus direction perception mechanism. The system automatically assigns fusion weights based on the difference in depth of the observed focus and the depth of each voxel layer, ensuring sharpness in the focused layer and smooth attenuation in the out-of-focus layer, thereby enhancing the sense of spatial depth.

[0098] (1) Set the focus and depth parameters

[0099] Let the current focus depth be The center depth of each voxel layer is ( The system determines the sharpness contribution of each layer by calculating the depth difference between different layers and the focal point.

[0100] (2) Calculate the focus response weight

[0101] A Gaussian focus response function is constructed based on the focusing behavior of the human eye, and a depth gradient term is introduced on this basis to reflect the focal length sensitivity of the scene edge.

[0102] (3) Perform voxel focusing fusion

[0103] The multi-voxel results are weighted and fused according to the focus response weight, so that the features at the focus point are enhanced and the features at the defocus point are naturally attenuated, so as to obtain a focusing effect consistent with human eye perception.

[0104] Focus weighting function:

[0105]

[0106] in: Indicates the first The focus response weight of the layer. This is a normalization constant used to ensure that all weights satisfy... . For the first Depth at the center of the layer. This is the current focus depth. This represents the depth-of-field width parameter, which controls the range of the area in focus. This represents the influence coefficient of the depth gradient. The depth gradient magnitude represents the local depth change rate.

[0107] The function controls the depth of field range through a Gaussian term, making the in-focus area sharp and the out-of-focus area smoothly faded; the depth gradient term enhances the response at object boundaries and the foreground-background boundary, making the system more sensitive to changes in three-dimensional layering.

[0108]

[0109] in The result is a focus-weighted fusion. Indicates the first Weighted fusion image of layers. This is the focus response weight, used to control the contribution intensity of each layer. This represents the total number of voxel layers.

[0110] This formula uses focus-related weighted fusion to give greater weight to layers near the focus depth, while layers farther from the focus are smoothly attenuated. The fusion result is clear and sharp in the focused area and naturally blurred in the out-of-focus area, thus achieving a visual depth-of-field effect consistent with the focusing mechanism of the real human eye.

[0111] Technical effects: This step achieves physical simulation of the focusing characteristics of the human eye by constructing a focus perception weighted model; the focus area is significantly enhanced, and the defocus area transitions naturally; the depth of field distribution is continuous and smooth, and the visual layers are realistic; it effectively reduces the phenomena of "depth of field jump" and "discontinuity between reality and illusion" in traditional light field displays.

[0112] Step 4: Multi-domain fusion and adaptive compensation output

[0113] Explanation: In the process of multi-layer voxel image fusion, direct RGB overlay can easily lead to edge blurring and cross-layer error accumulation. To address this, this invention proposes a multi-domain fusion mechanism based on structural consistency and depth compensation, achieving improved geometric accuracy and detail fidelity of the image within a framework of joint optimization in the spatial and frequency domains. By introducing depth residual terms and spatial consistency constraints during the fusion stage, the system achieves smooth cross-layer transitions and overall brightness self-balancing, obtaining a natural visual effect without relying on a luminance-chrominance separation model.

[0114] (1) Frequency domain energy enhancement

[0115] The fusion result By converting to the frequency domain, low-frequency blurring is suppressed and high-frequency detail components are amplified through energy enhancement operations to restore structural texture and edge sharpness;

[0116] (2) Depth residual compensation

[0117] Based on frequency domain optimization, a depth residual term is introduced to correct the geometric deviation between voxel layers, thereby improving depth continuity and spatial alignment accuracy.

[0118] (3) Gradient Consistency Optimization

[0119] Gradient consistency constraints are applied in the spatial domain to adaptively adjust brightness and structure, ensuring smooth edge transitions and maintaining global illumination balance.

[0120] Fusion expression:

[0121]

[0122] in: This represents the Fourier transform. For depth Laplace terms. This is the structural compensation coefficient. This indicates the final output image. These represent the Fourier transform and the inverse transform, respectively. The Laplacian term of the depth map is used to enhance the edges of structures. This is the structural compensation coefficient, used to control the depth compensation intensity.

[0123] This formula represents structural enhancement based on depth information in the frequency domain, using a multiplicative compensation factor. High-frequency details are enhanced, and then restored to the spatial domain through inverse transformation, achieving synergistic optimization of detail enhancement and illumination balance.

[0124] Technical effects: Through multi-domain fusion and adaptive compensation mechanisms, this step achieves joint optimization of depth information and brightness structure; effectively suppresses cross-layer artifacts and illumination discontinuities; significantly enhances the detail level and depth consistency of the output image, and improves the natural sense of layering and high-fidelity display quality of the 3D reconstruction results.

[0125] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions that fall within the scope of the present invention are within the scope of protection of the present invention.

Claims

1. A multi-domain voxel reconstruction three-dimensional display method, characterized in that: The following steps are included: S1: Voxel layer feature extraction and convolutional reconstruction: The input voxel layer is used to extract features and repair the structure using a convolutional neural network. Specifically, the following steps are included: (1) Input voxel sequence: Input a multi-plane image voxel sequence I(x, y, l), where l ∈ [1, 32] represents the depth layer index, each layer corresponds to a different spatial depth slice in the scene, and x, y represent the horizontal and vertical coordinates of the plane image; (2) Multi-branch convolutional feature extraction: Each branch is set with a convolutional kernel of different scales. (2) Used to extract local texture and global geometric features, and establish cross-depth dependencies between layers; (3) In the feature fusion stage, a two-dimensional Laplacian operator is introduced. The formula used to enhance the feature gradient is: ,in This represents the input voxel feature map of the l-th layer, where r is the radius of the neighboring depth layer. This represents the weighted convolution operation of the i-th scale convolution branch on the features of adjacent layers. This represents the second-order spatial Laplacian enhancement of the features of this layer. The structural reinforcement factor is... (4) Perform feature backpropagation and cross-layer reconstruction: Introduce residual units to realize cross-layer feature feedback and upsampling repair: ,in The output features are reconstructed for the l-th layer. This indicates an upsampling operation, which modifies the features from the previous layer... Interpolate to the current resolution. This represents the cross-layer fusion balance factor, used to adjust the contribution of inter-layer residuals. S1: Represents the feature residual between this layer and the previous layer, used to capture the subtle differences caused by depth variations; S2: Voxel layer fusion based on learned weights: Achieves natural inter-layer transitions and structural enhancement through learned weight allocation and edge compensation mechanisms; S3: Perceptual fusion weight modeling in the focal direction: Based on the difference between the observed focal depth and the depth of each voxel layer, the fusion weight is automatically assigned to make the focused layer clear and the defocused layer smoothly attenuated, thereby enhancing the spatial stereoscopic effect. Specifically, it includes the following steps: (1) Setting the focal and layer depth parameters: Let the current focal depth be 1. The central depth of each voxel layer is The system determines the sharpness contribution of each layer by calculating the depth difference between different layers and the focus. (2) Calculate the focus response weights : Construct a Gaussian focus response function based on human eye focusing behavior, and introduce a depth gradient term to reflect the focal length sensitivity of scene edges; (3) Perform voxel focus fusion: Perform weighted fusion of multi-layer voxel results according to focus response weights, so that the features at the focus point are enhanced and the defocus points are naturally attenuated, so as to obtain a focusing effect consistent with human eye perception, where the focus weighting function is: in Indicates the first Layer focus response weights, The normalization constant is For the first Layer center depth, This represents the depth of field width parameter. The depth gradient influence coefficient. The depth gradient magnitude represents the local depth change rate. The focus-weighted fusion result is as follows: in The result of focus-weighted fusion. Indicates the first Layer-weighted fusion image, To indicate the first Layer focus response weight; S4: Image output after multi-domain fusion and adaptive compensation: In the fusion stage, a depth residual term and spatial consistency constraint are introduced to achieve smooth transition across layers and overall brightness self-balancing.

2. The multi-domain voxel reconstruction three-dimensional display method according to claim 1, characterized in that: The establishment of cross-depth dependency in step (2) of S1 includes the following steps: by feeding adjacent depth layers together into the attention matrix, the features of each layer not only depend on itself, but also on other depth layers, as follows: In the deep attention scheme, the input voxel I(x,y,l) of each depth layer first passes through the convolution kernel W i Extract local geometric features to obtain the corresponding query vector Q. l Key vector K l AND value vector V l Then the attention matrix A(l,k) is based on Q l and K k The similarity learning determines the degree of dependency between different depth layers, and this is used to apply the value vector V of different depth layers. k Weighted fusion is performed to establish long-range cross-depth feature dependencies, where the attention matrix expression is as follows: Where l and k are depth layer indices, and k is the other depth layer that is associated with the l-th layer for calculation; The final generated layer l fusion feature for: Convolution kernels generate Q features Used to measure the depth structure of this layer: Convolution kernel generates K features Used to characterize the geometric features that can be provided: The convolutional kernel generates V features, which are used for the feature output after cross-depth fusion. convolution kernel All are weighted parameters. The transpose of the query vector is used to calculate cross-depth feature similarity, and Softmax() normalizes the similarity across all depth layers into attention weights.

3. The multi-domain voxel reconstruction three-dimensional display method according to claim 1, characterized in that: Step S2 includes the following steps: (1) Calculate the fusion weights for each layer : Targeting pixels At the given location, the learned weight network adaptively estimates the fusion weights for layers l at different depths. This is to reflect the credibility and salience of each layer in the local area; (2) Weighting the multi-layer voxel results by weight: Based on the voxel reconstruction results of each layer... With weight Perform a weighted average operation to generate a fused image; (3) Edge-guided blur region restoration: In the fused image, edge-guided items are used... Enhancement and repair of blurred areas are performed by strengthening edge energy response through inverse frequency domain transform, restoring structural clarity and contour sharpness. The core expression is: in To produce the merged output image, This is the result of voxel reconstruction of layer l. These are the fusion weights, reflecting the contribution of this layer at pixel p. It is the inverse Fourier transform, used to recover edge energies in the frequency domain. The gradient distribution of the edge map, is the edge enhancement coefficient, which controls the repair intensity, and L is the total number of layers.

4. The multi-domain voxel reconstruction three-dimensional display method according to claim 1, characterized in that: Step S4 includes the following steps: (1) Frequency domain energy enhancement: the fusion result Convert to frequency domain: Inverse transformation Energy-enhanced operations suppress low-frequency blurring and amplify high-frequency detail components to restore structural texture and edge sharpness. This represents the focus-weighted fused image. express Two-dimensional discrete Fourier transform, For pixel coordinates, u and v are frequency domain indices or normalized frequencies; (2) Depth residual compensation: Based on frequency domain optimization, a depth residual term is introduced to correct the geometric deviation between voxel layers, thereby improving depth continuity and spatial alignment accuracy. (3) Gradient consistency optimization: Apply gradient consistency constraints in the spatial domain to adaptively adjust brightness and structure, ensuring smooth edge transition and maintaining global illumination balance; (4) Image output: in: Indicates Fourier transform, For the depth Laplace term, This is the structural compensation coefficient. This indicates the final output image. These represent the Fourier transform and the inverse transform, respectively. The Laplacian term of the depth map is used to enhance the edges of structures.

5. A multi-domain voxel reconstruction three-dimensional display device, characterized in that, It is displayed using the three-dimensional display method as described in any one of claims 1-4.

Images