Strong interference environment holographic enhancement system and method based on reverse diffusion generation

Abstract

According to the strong interference environment holographic enhancement system and method based on reverse diffusion generation, an image set in a strong interference environment is collected, and a static body which does not change drastically along with time and a dynamic body which flows at high frequency are identified according to static images and dynamic images in the image set; the method comprises the following steps: constructing a double-flow implicit field comprising a static geometric field and a dynamic medium field, performing double-flow implicit field body rendering on a static body and a dynamic body, classifying the static body into the static geometric field in a density and color field, classifying the dynamic body into the dynamic medium field in the density and color field, obtaining implicit features, and enabling the implicit features to adapt to a complex scene of a strong interference environment. High-quality two-dimensional visual input is provided for subsequent three-dimensional reconstruction, denoising sampling is carried out on implicit features by using a physically guided reverse diffusion model, multi-physics field constraint loss is introduced in the model training process, the reliability is quantified according to the implicit feature decomposition cognitive uncertainty and accidental uncertainty, and then the implicit features are extracted. And a reconstruction result of the holographic three-dimensional scene is generated, model inknowledge and data noise are distinguished, and the reconstruction precision of the holographic three-dimensional scene is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of holographic enhancement technology for strong interference environments, and specifically to a holographic enhancement system and method for strong interference environments based on reverse diffusion generation. Background Technology

[0002] Holographic 3D reconstruction and visual enhancement in environments with strong interference is a cutting-edge technology field at the intersection of industrial environmental perception and computer vision. It focuses on achieving high-precision holographic reconstruction and visual enhancement of 3D scenes in complex industrial aerosol interference scenarios such as dust, water mist, and dense smoke. At the same time, it provides a reliable quantitative evaluation of the reconstruction results, providing technical support for perception and decision-making in scenarios such as industrial automation, emergency rescue, and security monitoring.

[0003] In existing industrial scenarios, dust and water mist often exhibit non-uniform and high-concentration distribution characteristics. Fixed physical model parameters are difficult to adapt to complex scenarios, and problems such as incomplete defogging, image color distortion, and incorrect dynamic medium morphology restoration are prone to occur. This makes it impossible to provide high-quality 2D visual input for subsequent 3D reconstruction. Moreover, existing 3D reconstruction algorithms only focus on the restoration of scene geometry and visual effects, without quantitatively evaluating the uncertainty of the reconstruction results point by point, which greatly reduces the accuracy of subsequent holographic reconstruction. Summary of the Invention

[0004] The present invention proposes a holographic enhancement system and method for strong interference environments based on reverse diffusion generation. This system addresses the challenges in existing industrial scenarios where dust and water mist often exhibit non-uniform, high-concentration distributions. Fixed physical model parameters are difficult to adapt to complex scenes, leading to problems such as incomplete defogging, image color distortion, and errors in the restoration of dynamic medium morphology. These issues prevent the provision of high-quality 2D visual input for subsequent 3D reconstruction. Furthermore, existing 3D reconstruction algorithms only focus on restoring the scene's geometry and visual effects, without quantitatively evaluating the uncertainty of the reconstruction results point by point, resulting in a significant reduction in the accuracy of subsequent holographic reconstruction.

[0005] To achieve the above objectives, the present invention employs the following technical solution: The strong interference environment holographic enhancement system based on reverse diffusion generation of the present invention comprises: The image acquisition module is used to acquire an image set in a strong interference environment and output the image set, which includes static images at various time points and dynamic images over a time period. An image recognition module is used to acquire the image set and, based on the static and dynamic images in the image set, identify static bodies that do not change drastically over time and dynamic bodies that flow at high frequencies. The module constructs a dual-flow implicit field containing a static geometric field and a dynamic medium field, and performs dual-flow implicit field volume rendering on the static and dynamic volumes. The static and dynamic volumes in the three-dimensional space are mapped to continuous density and color fields. The static volumes are assigned to the static geometric field in the density and color fields, and the dynamic volumes are assigned to the dynamic medium field in the density and color fields, thus deriving implicit features. The processing module is used to perform noise reduction sampling on implicit features using a physics-guided reverse diffusion model to remove environmental noise from the implicit features. The physical constraint optimization module is used to introduce multi-physics constraint loss during model training so that implicit features satisfy the laws of momentum conservation and viscous dissipation. The holographic evaluation generation module is used to quantify credibility by decomposing cognitive uncertainty and accidental uncertainty based on implicit features, and to generate the reconstruction results of the holographic 3D scene.

[0006] Preferably, the building module includes: The dual-stream data adaptation and sampling submodule is used to convert the recognition results of static and dynamic objects into structured sampling data required for implicit field modeling, and project the three-dimensional implicit field into a two-dimensional image through integral equations. The static geometric field modeling submodule is used to learn the continuous implicit representation of a static volume and outputs the density field and color field of the static volume. The dynamic medium field modeling submodule is used to learn the temporal continuous implicit representation of a dynamic volume and output the density field and color field of the dynamic volume. The dual-flow-field collaborative optimization submodule is used to integrate the static geometric field and the dynamic medium field to resolve spatial overlap conflicts; The implicit feature extraction and encoding submodule is used to extract core features from the optimized two-stream implicit field to form implicit features; The formula for calculating the integral equation is as follows: ; ; in, The color value of the two-dimensional pixel corresponding to the camera's line of sight. Near-plane depth boundary for volume rendering. The far-plane depth boundary for volume rendering. Cumulative transmittance For static bodies in the radiation depth Volume density at that location The line-of-sight ray emitted by the camera. The extinction coefficient of the medium, For dynamic bodies in rays depth Volume density at that location The color radiance is the result of fusing the static geometric field and the dynamic medium field. For media validity mask, This is a multi-view feature aggregation vector. Perspective weight; ; in, .

[0007] Preferably, the formula for calculating the latent variables of the implicit features in backdiffusion is as follows: ; in, Reverse diffusion process Hidden variables of the step, Reverse diffusion process Hidden variables of the step, These are the preset noise scheduling parameters. Cumulative noise scheduling parameters For parameters The noise component predicted by a denoising network that incorporates optical flow features. The variance matrix predicted by the model. The fluid energy potential function is defined. For the first The noise standard deviation of the diffusion process. For random noise that follows a Gaussian distribution, For time-series smoothing energy potential function, This represents the timing constraint coefficient.

[0008] Preferably, the holographic evaluation generation module includes: The implicit feature adaptation preprocessing submodule is used to standardize the input implicit features; The dual-type uncertainty decomposition submodule is used to calculate the cognitive uncertainty and accidental uncertainty of holography based on deep integration and variational inference, respectively. The holographic 3D reconstruction result generation submodule is used to fuse implicit features and credibility data to generate holographic 3D scene reconstruction results with credibility labels. The reconstruction result verification and optimization submodule is used to verify the consistency of the generated holographic reconstruction results and credibility data, and to eliminate artifacts and calculation errors.

[0009] Preferably, the formula for calculating holographic uncertainty is as follows: ; in, The total uncertainty of a holographic 3D scene, This represents the number of sub-models in a deep ensemble model. For the first The average density or color prediction of each sub-model for a 3D scene. For all The average of the predicted means of each sub-model For the first The variance of the prediction results of each sub-model for the 3D scene. For spatial correlation weights, For neighborhood uncertainty correlation regularization term, The correlation regularization coefficient is used. To understand uncertainty, It is due to chance and uncertainty.

[0010] Preferably, the physical constraint optimization module includes: The photometric reconstruction error submodule is used to measure the photometric difference between the rendered image and the actual observed image; The fluid divergence constraint submodule is used to penalize non-physical compression or expansion of the medium field, so that the diffusion of smoke and dust conforms to the fluid continuity equation; The curl consistency constraint submodule is used to ensure that the local rotational characteristics of the flow field conform to the statistical laws of the prior vorticity field. The computation module is used to calculate the total loss function for model training based on implicit features.

[0011] Preferably, the formula for calculating the total loss function during model training is as follows: ; in, The total loss function for model training. For photometric reconstruction error, This is a spatially adaptive prior vorticity field. The weighting coefficients for the fluid divergence constraint loss are... Dynamic medium field velocity vector field divergence, The weighting coefficients for the curl consistency constraint loss are: For dynamic medium field velocity vector field curl, For voxel confidence weights, The true value of the static geometric normal vector. These are the geometric constraint coefficients.

[0012] Preferably, the image acquisition module includes: The first acquisition submodule is used to acquire static images at various time points in a highly interference environment; The second acquisition submodule is used to acquire dynamic images over a period of time in a highly interference environment; The output module is used to output static images at various time points and dynamic images over a period of time.

[0013] Preferably, the image recognition module includes: The time dimension feature extraction submodule is used to capture the change information of the image set on the time axis and quantify the core features of static and dynamic data. The static / dynamic feature separation submodule is used to perform semantic recognition on the separated features and clearly label the spatial location and attributes of static and dynamic objects. The temporal consistency verification submodule is used for multi-frame cross-verification of target types to avoid misjudgment in a single frame and ensure the temporal stability of the recognition results.

[0014] The method for a holographic enhancement system for a strongly interfering environment based on reverse diffusion includes the following steps: S1. Acquire an image set in a strong interference environment and output the image set, which includes static images at various time points and dynamic images over a time period; S2. Obtain the image set, and based on the static and dynamic images in the image set, identify static bodies that do not change drastically over time and dynamic bodies that flow at high frequency. S3. Construct a dual-flow implicit field containing a static geometric field and a dynamic medium field, and perform dual-flow implicit field volume rendering on the static and dynamic volumes. Map the static and dynamic volumes in the three-dimensional space to continuous density and color fields. Substitute the static volume into the static geometric field in the density and color field, and the dynamic volume into the dynamic medium field in the density and color field to obtain implicit features. S4. Using a physics-guided reverse diffusion model, denoise sampling is performed on the implicit features to remove environmental noise from the implicit features. S5. Introduce multiphysics constraint loss during model training to make implicit features satisfy the laws of momentum conservation and viscous dissipation. S6. Based on the implicit feature decomposition of cognitive uncertainty and accidental uncertainty, the credibility is quantified, and the reconstruction result of the holographic 3D scene is generated.

[0015] As can be seen from the above technical solution, the present invention provides a holographic enhancement system for strong interference environments based on reverse diffusion generation. Compared with the prior art, the present invention has the following advantages: By acquiring an image set in a strong interference environment, and based on the static and dynamic images in the image set, static bodies that do not change drastically over time and dynamic bodies that flow at high frequencies are identified. A dual-flow implicit field containing a static geometric field and a dynamic medium field is constructed, and the static and dynamic bodies are rendered using the dual-flow implicit field. The static bodies are classified into the static geometric field in the density and color fields, and the dynamic bodies are classified into the dynamic medium field in the density and color fields, thus obtaining implicit features that are adapted to the complex scenes of strong interference environments, providing high-quality 2D visual input for subsequent 3D reconstruction. Using a physics-guided reverse diffusion model, the implicit features are denoised and sampled. Multiphysics constraint loss is introduced during model training. Based on the implicit features, cognitive uncertainty and accidental uncertainty are decomposed to quantify the credibility, generating the reconstruction result of the holographic 3D scene. The model ignorance and data noise are distinguished, greatly improving the reconstruction accuracy of the holographic 3D scene. Attached Figure Description

[0016] Figure 1 This is a structural block diagram of the holographic enhancement system for strong interference environments based on reverse diffusion generated according to the present invention; Figure 2 This is a flowchart illustrating the holographic enhancement method for strong interference environments based on reverse diffusion generated according to the present invention. Figure 3 This is a schematic diagram of the uncertainty assessment process of the present invention.

[0017] The attached figures are labeled as follows: Image acquisition module 111; image recognition module 112; construction module 113; processing module 114; physical constraint optimization module 115; holographic evaluation generation module 116. Detailed Implementation

[0018] 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 some embodiments of the present invention, but not all embodiments.

[0019] like Figure 1 As shown, the holographic enhancement system for strong interference environments based on reverse diffusion generation in this embodiment includes: The image acquisition module is used to acquire and output an image set in a highly interference environment. The image set includes static images at various time points and dynamic images over a period of time. The image recognition module is used to acquire an image set and, based on the static and dynamic images in the image set, identify static bodies that do not change drastically over time and dynamic bodies that flow at high frequencies. The module constructs a dual-flow implicit field containing a static geometric field and a dynamic medium field, and performs dual-flow implicit field volume rendering on static and dynamic volumes. It maps static and dynamic volumes in three-dimensional space to continuous density and color fields, classifies static volumes into the static geometric field in the density and color field, and classifies dynamic volumes into the dynamic medium field in the density and color field, thus deriving implicit features. The processing module is used to perform noise reduction sampling on implicit features using a physics-guided reverse diffusion model to remove environmental noise from the implicit features. The physical constraint optimization module is used to introduce multi-physics constraint loss during model training so that implicit features satisfy the laws of momentum conservation and viscous dissipation. The holographic evaluation generation module is used to quantify credibility by decomposing cognitive uncertainty and accidental uncertainty based on implicit features, and to generate the reconstruction results of the holographic 3D scene.

[0020] The building blocks include: The dual-stream data adaptation and sampling submodule is used to convert the recognition results of static and dynamic objects into structured sampling data required for implicit field modeling, and project the three-dimensional implicit field into a two-dimensional image through integral equations. The static geometric field modeling submodule is used to learn the continuous implicit representation of a static volume and outputs the density field and color field of the static volume. The dynamic medium field modeling submodule is used to learn the temporal continuous implicit representation of a dynamic volume and output the density field and color field of the dynamic volume. The dual-flow-field collaborative optimization submodule is used to integrate the static geometric field and the dynamic medium field to resolve spatial overlap conflicts; The implicit feature extraction and encoding submodule is used to extract core features from the optimized two-stream implicit field to form implicit features; The formula for calculating the integral equation is as follows: ; ; in, The color value of the two-dimensional pixel corresponding to the camera's line of sight. Near-plane depth boundary for volume rendering. The far-plane depth boundary for volume rendering. Cumulative transmittance For static bodies in the radiation depth Volume density at that location The line-of-sight ray emitted by the camera. The extinction coefficient of the medium, For dynamic bodies in rays depth Volume density at that location The color radiance is the result of fusing the static geometric field and the dynamic medium field. For media validity mask, This is a multi-view feature aggregation vector. Perspective weight; ; in, .

[0021] In real-world fire rescue scenes, dense smoke obscures internal equipment. It's necessary to first decouple static elements (walls and equipment, etc.) from dynamic smoke using volumetric rendering to generate a preliminary 3D visualization image. Then, a rescue robot carrying a camera enters the fire scene. At this point, the camera... Get the camera coordinates of the rescue robot (0 0 1.2 ), the robot's line of sight Pointing towards the interior of the fire (0.5) 0.3 —0.8 ), near-plane depth boundary of volume rendering The far-plane depth boundary of the volume rendering It can cover 5 fire areas In the core area, the empirical value for static volume density is 1.2; higher values ​​indicate opaque rigid bodies, walls, or equipment. , The value is 0.7 in the core area of ​​the dense smoke and 0.2 in the edge area; the medium extinction coefficient The core area is predicted by a neural network based on the smoke concentration. =0.9 represents strong occlusion, in the edge area. This represents weak occlusion; Calculate Integrating depth features from three perspectives If this feature is not present from a single viewpoint, a wall ghost image will appear. At this moment, the line of sight for ; Pick arrive ; ; exist At that time, calculate That is, the light reaches 3 m At depth, only 1% was not obscured by dense smoke; calculate When integrating over the depth interval, static walls or equipment Depend on Calibration was performed, correcting the wall color from (0.15, 0.15, 0.15) from a single viewpoint to (0.4, 0.4, 0.4) from a multi-viewpoint fusion. Near-view weight Far-view weight Prioritize the use of clear features from close-up views, and weight the color values ​​to (0.4, 0.4, 0.4). = (0.36, 0.36, 0.36); Calculation The value is (0.21, 0.21, 0.21), which means that the wall or equipment is obscured by thick smoke, and the original color is weakened.

[0022] This way and Distinguish between walls and dense smoke to prevent dense smoke from being mistaken for a static obstacle, and By dynamically adjusting transmittance based on smoke concentration, reducing wall visibility in the core obstruction zone, and using a medium effectiveness mask, the effectiveness of the medium is effectively masked. After eliminating dust and noise, the accuracy of dynamic media recognition improved from 72% to 91%, thanks to multi-view features. Eliminating single-view ghosting, the PSNR value of the rendered image is increased from 22. dB Upgraded to 28 dB .

[0023] The formula for calculating the latent variables of implicit features in backdiffusion is as follows: ; in, Reverse diffusion process Hidden variables of the step, Reverse diffusion process Hidden variables of the step, These are the preset noise scheduling parameters. For cumulative noise scheduling parameters, For parameters The noise component predicted by a denoising network that incorporates optical flow features. The variance matrix predicted by the model. The fluid energy potential function is defined. For the first The noise standard deviation of the diffusion process. For random noise that follows a Gaussian distribution, For time-series smoothing energy potential function, These are the timing constraint coefficients. The physical guiding strength coefficient. This is the variance dynamic scaling factor.

[0024] The implicit features are severely obscured by dense smoke, so reverse diffusion is needed to denoise them. At the same time, it is necessary to ensure that the flow of dense smoke conforms to the fluid laws and avoid generating voids or artifacts without physical logic. In practical applications, reverse diffusion takes 800 steps, that is... At this time, noise scheduling parameters Cumulative noise scheduling parameters Physical guiding strength coefficient Noise standard deviation , ; During calculation, Input from a 3D U-Net network Noisy latent variables of the step Output noise ; calculate This item is used for initial removal of Gaussian noise; Variance matrix of model prediction Set to 0.02, current gradient A value of 0.3 forces the smoke flow field to correct in a direction that conforms to aerodynamics. The concentration of dense smoke is dynamically determined, with a core area concentration of 15. Maximum concentration 20 Therefore The physical correction term is ; This represents the temporal deviation between the current latent variable and the previous latent variable; the larger the deviation, the higher the gradient value. and If the flow velocity deviation is 0.05, then , =0.25; The decay effect decreases with each diffusion step; physical constraints are strengthened in the early stages, while temporal smoothing is strengthened in the later stages. hour, , thus calculating ; Variance dynamic scaling factor Based on uncertainty calculations, in a dense smoke environment, the following results are obtained: , ; Calculation Hidden variables of time ; Thus, through Temporal smoothing constraints reduce the inter-frame jump amplitude of the dense smoke flow field from that in traditional formulas. Down to The timing consistency compliance rate increased from 68% to 95%; fusion After analyzing the optical flow temporal characteristics, the noise prediction error of the denoising network decreases from... Down to The clarity of the smoke outline improved by 32% after single-step noise reduction; Differentiated physical guidance is achieved for the core and edge regions. The physical constraint intensity in the core region (high concentration) is increased by 50%, avoiding "smoke backflow" artifacts; the constraint in the edge region (low concentration) is reduced, and the detail retention rate is increased from 62% to 89%. The holographic evaluation generation module includes: The implicit feature adaptation preprocessing submodule is used to standardize the input implicit features; The dual-type uncertainty decomposition submodule is used to calculate the cognitive uncertainty and accidental uncertainty of holography based on deep integration and variational inference, respectively. The holographic 3D reconstruction result generation submodule is used to fuse implicit features and credibility data to generate holographic 3D scene reconstruction results with credibility labels. The reconstruction result verification and optimization submodule is used to verify the consistency of the generated holographic reconstruction results and credibility data, and to eliminate artifacts and calculation errors.

[0025] Its key feature is that the formula for calculating holographic uncertainty is as follows: ; in, The total uncertainty of a holographic 3D scene, This represents the number of sub-models in a deep ensemble model. For the first The average density or color prediction of each sub-model for a 3D scene. For all The average of the predicted means of each sub-model For the first The variance of the prediction results of each sub-model for the 3D scene. For spatial correlation weights, For neighborhood uncertainty correlation regularization term, The correlation regularization coefficient is used. To understand uncertainty, It is due to chance and uncertainty; In high-risk environments such as mines or dusty workshops, AGVs need to avoid high-risk areas based on the reliability of the reconstructed point cloud. like Figure 3 As shown, taking a dust workshop as an example, uncertainty assessment is performed, and the number of deeply integrated models is [number missing]. ,Right now Voxel to be tested For the corner of the workshop This area is a blind spot for equipment, resulting in limited observation. Five sub-models for voxels Density prediction mean They are respectively Prediction variance They are respectively , ; Mean variance of neighborhood of dead zone global ,but Mean variance of neighborhood in open area , ; Cognitive uncertainty The low spatial correlation in blind spots highlights cognitive ignorance when weighted. Random uncertainty The random uncertainty is negative, so it should be appropriately reduced to highlight the dominance of cognitive uncertainty. It is 0.00144; Total uncertainty ; Spatial correlation weight By highlighting the uncertainty in blind spots, AGVs can improve decision-making accuracy by 30% by bypassing them.

[0026] The physical constraint optimization module includes: The photometric reconstruction error submodule is used to measure the photometric difference between the rendered image and the actual observed image; The fluid divergence constraint submodule is used to penalize non-physical compression or expansion of the medium field, so that the diffusion of smoke and dust conforms to the fluid continuity equation; The curl consistency constraint submodule is used to ensure that the local rotational characteristics of the flow field conform to the statistical laws of the prior vorticity field. The computation module is used to calculate the total loss function for model training based on implicit features.

[0027] The formula for calculating the total loss function during model training is as follows: ; in, The total loss function for model training. For photometric reconstruction error, This is a spatially adaptive prior vorticity field. The weighting coefficients for the fluid divergence constraint loss are... Dynamic medium field velocity vector field divergence, The weighting coefficients for the curl consistency constraint loss are: For dynamic medium field velocity vector field curl, For voxel confidence weights, The true value of the static geometric normal vector. These are the geometric constraint coefficients; In actual use, after noise reduction, it was found that there were "high-frequency jumps" in local dense smoke areas (such as a sudden increase in the density of dense smoke in a certain area). It is necessary to optimize the network parameters through physical constraint loss to correct non-physical artifacts and luminance.

[0028] The image acquisition module includes: The first acquisition submodule is used to acquire static images at various time points in a highly interference environment; The second acquisition submodule is used to acquire dynamic images over a period of time in a highly interference environment; The output module is used to output static images at various time points and dynamic images over a period of time.

[0029] The image recognition module includes: The time dimension feature extraction submodule is used to capture the change information of the image set on the time axis and quantify the core features of static and dynamic data. The static / dynamic feature separation submodule is used to perform semantic recognition on the separated features and clearly label the spatial location and attributes of static and dynamic objects. The temporal consistency verification submodule is used for multi-frame cross-verification of target types to avoid misjudgment in a single frame and ensure the temporal stability of the recognition results.

[0030] like Figure 2 As shown, the method for a holographic enhancement system for strong interference environments based on reverse diffusion includes the following steps: S1. Acquire an image set in a strong interference environment and output the image set, which includes static images at various time points and dynamic images over a period of time. S2. Obtain the image set, and based on the static and dynamic images in the image set, identify static bodies that do not change drastically over time and dynamic bodies that flow at high frequency. S3. Construct a dual-flow implicit field containing a static geometric field and a dynamic medium field, and perform dual-flow implicit field volume rendering on static and dynamic volumes. Map static and dynamic volumes in three-dimensional space to continuous density and color fields. Substitute static volumes into the static geometric field in the density and color field, and dynamic volumes into the dynamic medium field in the density and color field to obtain implicit features. S4. Using a physics-guided reverse diffusion model, denoise sampling is performed on the implicit features to remove environmental noise from the implicit features. S5. Introduce multiphysics constraint loss during model training to make implicit features satisfy the laws of momentum conservation and viscous dissipation. S6. Based on the implicit feature decomposition of cognitive uncertainty and accidental uncertainty, the credibility is quantified, and the reconstruction result of the holographic 3D scene is generated.

[0031] As can be seen from the above technical solution, this invention provides a holographic enhancement system for strong interference environments based on reverse diffusion generation. Compared with existing technologies, this invention has the following advantages: By acquiring an image set in a strong interference environment, and based on the static and dynamic images in the image set, static bodies that do not change drastically over time and dynamic bodies that flow at high frequencies are identified. A dual-flow implicit field containing a static geometric field and a dynamic medium field is constructed, and the static and dynamic bodies are rendered using the dual-flow implicit field. The static bodies are classified into the static geometric field in the density and color fields, and the dynamic bodies are classified into the dynamic medium field in the density and color fields, thus deriving implicit features that adapt to the complex scenes of strong interference environments, providing high-quality 2D visual input for subsequent 3D reconstruction. Using a physics-guided reverse diffusion model, the implicit features are denoised and sampled. Multiphysics constraint loss is introduced during model training. Based on the implicit features, cognitive uncertainty and accidental uncertainty are decomposed to quantify credibility, generating holographic 3D scene reconstruction results. This distinguishes between model ignorance and data noise, greatly improving the reconstruction accuracy of holographic 3D scenes. Therefore, this invention effectively overcomes the various shortcomings of existing technologies and has high industrial application value.

[0032] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state disk (SSD)).

[0033] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0034] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0035] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. The holographic enhancement system for strong interference environments based on reverse diffusion generation of the present invention is characterized in that, include: The image acquisition module is used to acquire an image set in a strong interference environment and output the image set, which includes static images at various time points and dynamic images over a time period. An image recognition module is used to acquire the image set and, based on the static and dynamic images in the image set, identify static bodies that do not change drastically over time and dynamic bodies that flow at high frequencies. The module constructs a dual-flow implicit field containing a static geometric field and a dynamic medium field, and performs dual-flow implicit field volume rendering on the static and dynamic volumes. The static and dynamic volumes in the three-dimensional space are mapped to continuous density and color fields. The static volumes are assigned to the static geometric field in the density and color fields, and the dynamic volumes are assigned to the dynamic medium field in the density and color fields, thus deriving implicit features. The processing module is used to perform noise reduction sampling on implicit features using a physics-guided reverse diffusion model to remove environmental noise from the implicit features. The physical constraint optimization module is used to introduce multi-physics constraint loss during model training so that implicit features satisfy the laws of momentum conservation and viscous dissipation. The holographic evaluation generation module is used to quantify credibility by decomposing cognitive uncertainty and accidental uncertainty based on implicit features, and to generate the reconstruction results of the holographic 3D scene.

2. The holographic enhancement system for strong interference environments based on reverse diffusion generation according to claim 1, characterized in that: The building module includes: The dual-stream data adaptation and sampling submodule is used to convert the recognition results of static and dynamic objects into structured sampling data required for implicit field modeling, and project the three-dimensional implicit field into a two-dimensional image through integral equations. The static geometric field modeling submodule is used to learn the continuous implicit representation of a static volume and outputs the density field and color field of the static volume. The dynamic medium field modeling submodule is used to learn the temporal continuous implicit representation of a dynamic volume and output the density field and color field of the dynamic volume. The dual-flow-field collaborative optimization submodule is used to integrate the static geometric field and the dynamic medium field to resolve spatial overlap conflicts; The implicit feature extraction and encoding submodule is used to extract core features from the optimized two-stream implicit field to form implicit features; The formula for calculating the integral equation is as follows: ； ； in, The color value of the two-dimensional pixel corresponding to the camera's line of sight. Near-plane depth boundary for volume rendering. The far-plane depth boundary for volume rendering. Cumulative transmittance For static bodies in the radiation depth Volume density at that location The line-of-sight ray emitted by the camera. The extinction coefficient of the medium, For dynamic bodies in rays depth Volume density at that location The color radiance is the result of fusing the static geometric field and the dynamic medium field. For media validity mask, This is a multi-view feature aggregation vector. Perspective weight; ； in, .

3. The holographic enhancement system for strong interference environments based on reverse diffusion generation according to claim 1, characterized in that: The formula for calculating the latent variables of the implicit features in back diffusion is as follows: ； in, Reverse diffusion process Hidden variables of the step, Reverse diffusion process Hidden variables of the step, These are the preset noise scheduling parameters. Cumulative noise scheduling parameters For parameters The noise component predicted by a denoising network that incorporates optical flow features. The variance matrix predicted by the model. The fluid energy potential function is defined. For the first The noise standard deviation of the diffusion process. For random noise that follows a Gaussian distribution, For time-series smoothing energy potential function, This represents the timing constraint coefficient.

4. The holographic enhancement system for strong interference environments based on reverse diffusion generation according to claim 1, characterized in that: The holographic evaluation generation module includes: The implicit feature adaptation preprocessing submodule is used to standardize the input implicit features; The dual-type uncertainty decomposition submodule is used to calculate the cognitive uncertainty and accidental uncertainty of holography based on deep integration and variational inference, respectively. The holographic 3D reconstruction result generation submodule is used to fuse implicit features and credibility data to generate holographic 3D scene reconstruction results with credibility labels; The reconstruction result verification and optimization submodule is used to verify the consistency of the generated holographic reconstruction results and credibility data, and to eliminate artifacts and calculation errors.

5. The holographic enhancement system for strong interference environments based on reverse diffusion generation according to claim 4, characterized in that: The formula for calculating the uncertainty of holography is as follows: ； in, The total uncertainty of a holographic 3D scene, This represents the number of sub-models in a deep ensemble model. For the first The average density or color prediction of each sub-model for a 3D scene. For all The average of the predicted means of each sub-model For the first The variance of the prediction results of each sub-model for the 3D scene. For spatial correlation weights, For neighborhood uncertainty correlation regularization term, The correlation regularization coefficient is used. To understand uncertainty, It is due to chance and uncertainty.

6. The holographic enhancement system for strong interference environments based on reverse diffusion generation according to claim 1, characterized in that: The physical constraint optimization module includes: The photometric reconstruction error submodule is used to measure the photometric difference between the rendered image and the actual observed image; The fluid divergence constraint submodule is used to penalize non-physical compression or expansion of the medium field, so that the diffusion of smoke and dust conforms to the fluid continuity equation; The curl consistency constraint submodule is used to ensure that the local rotational characteristics of the flow field conform to the statistical laws of the prior vorticity field. The computation module is used to calculate the total loss function for model training based on implicit features.

7. The holographic enhancement system for strong interference environments based on reverse diffusion generation according to claim 6, characterized in that: The formula for calculating the total loss function during model training is as follows: ； in, The total loss function for model training. For photometric reconstruction error, For spatial adaptive prior vorticity field. The weighting coefficients for the fluid divergence constraint loss are... Dynamic medium field velocity vector field divergence, The weighting coefficients for the curl consistency constraint loss are: For dynamic medium field velocity vector field curl, For voxel confidence weights, The true value of the static geometric normal vector. These are the geometric constraint coefficients.

8. The holographic enhancement system for strong interference environments based on reverse diffusion generation according to claim 1, characterized in that: The image acquisition module includes: The first acquisition submodule is used to acquire static images at various time points in a highly interference environment; The second acquisition submodule is used to acquire dynamic images over a period of time in a highly interference environment; The output module is used to output static images at various time points and dynamic images over a period of time.

9. The holographic enhancement system for strong interference environments based on reverse diffusion generation according to claim 1, characterized in that: The image recognition module includes: The time dimension feature extraction submodule is used to capture the change information of the image set on the time axis and quantify the core features of static and dynamic data. The static / dynamic feature separation submodule is used to perform semantic recognition on the separated features and clearly label the spatial location and attributes of static and dynamic objects. The temporal consistency verification submodule is used for multi-frame cross-verification of target types to avoid misjudgment in a single frame and ensure the temporal stability of the recognition results.

10. A method applied to the holographic enhancement system for strong interference environments based on reverse diffusion generation as described in claims 1-9, characterized in that, Includes the following steps: S1. Acquire an image set in a strong interference environment and output the image set, which includes static images at various time points and dynamic images over a time period; S2. Obtain the image set, and based on the static and dynamic images in the image set, identify static bodies that do not change drastically over time and dynamic bodies that flow at high frequency. S3. Construct a dual-flow implicit field containing a static geometric field and a dynamic medium field, and perform dual-flow implicit field volume rendering on the static and dynamic volumes. Map the static and dynamic volumes in the three-dimensional space to continuous density and color fields. Substitute the static volume into the static geometric field in the density and color field, and the dynamic volume into the dynamic medium field in the density and color field to obtain implicit features. S4. Using a physics-guided reverse diffusion model, denoise sampling is performed on the implicit features to remove environmental noise from the implicit features. S5. Introduce multiphysics constraint loss during model training to make implicit features satisfy the laws of momentum conservation and viscous dissipation. S6. Based on the implicit feature decomposition of cognitive uncertainty and accidental uncertainty, the credibility is quantified, and the reconstruction result of the holographic 3D scene is generated.

Images