A high-fidelity radiation field reconstruction method and system based on polarization normal estimation

Abstract

The present invention discloses a high-fidelity radiation field reconstruction method and system based on polarization normal estimation. The method of the present invention comprises extracting polarization degree and polarization angle from polarization image sequence, extracting position code from RGB image sequence; inputting multi-channel tensor composed of polarization image sequence, polarization degree, polarization angle and position code into trained attention mechanism image segmentation network to obtain the estimation result of polarization normal vector; extracting camera pose matrix and establishing sparse point cloud from RGB image sequence by using motion structure recovery algorithm SfM; and realizing high-fidelity radiation field reconstruction based on the established sparse point cloud by using improved DN‑Splatter algorithm guided by the estimation result of polarization normal vector. The present invention aims to accelerate the convergence process of three-dimensional Gaussian ellipsoid optimization, make the generated mesh surface smoother, and effectively solve the problems of slow convergence of existing radiation field reconstruction and uneven surface when generating mesh.

Description

Technical Field

[0001] The present invention belongs to the field of high-precision three-dimensional model reconstruction, and in particular relates to a high-fidelity radiation field reconstruction method and system based on polarization normal estimation. Background Art

[0002] Three-dimensional reconstruction is to reconstruct objects or scenes from a single or multiple two-dimensional images. It is an important research direction in the field of computer vision and has wide applications in medical diagnosis, cultural relic reconstruction, navigation and positioning. Traditional three-dimensional reconstruction is based on the principle of multi-view geometry and is achieved through multiple geometric solution steps such as camera parameter estimation, dense point cloud reconstruction and surface reconstruction. Drawing on the ability of humans to reconstruct and perceive three-dimensional scenes based on the brain rather than geometric solution, it naturally extends to the process of using deep learning methods to simulate the construction of three-dimensional space through two-dimensional images. Domestic research has a certain foundation, mainly focusing on improving the overall accuracy and rendering rate of three-dimensional reconstruction, but the three-dimensional Gaussian splashing algorithm has low accuracy in fitting the surface due to the lack of geometric constraints. It is mainly reflected in the defects of artifacts and blurring in the surface optimization process, poor fitting consistency, and slow convergence speed of complex scene optimization. Given that the polarization state of light on the surface of different objects is different, polarization surface shape measurement is also one of the specific applications of polarized light. The calculation of polarization normal vector based on physical models is difficult to overcome the noise interference caused by complex scenes and illumination, and there is no good solution to deal with the π ambiguity problem. With the development of deep learning technology, image segmentation networks have been widely used in polarization normal vector extraction. At present, the object-level polarization normal vector estimation has been extended to the scene level. However, the generalization of the network is not ideal, and it is easy to make incorrect estimates for complex scenes. Summary of the invention

[0003] Technical problem to be solved by the present invention: In view of the above-mentioned problems in the prior art, a high-fidelity radiation field reconstruction method and system based on polarization normal estimation are provided. The present invention aims to accelerate the convergence process of three-dimensional Gaussian ellipsoid optimization, make the generated mesh surface smoother, and effectively solve the problems of slow convergence of existing radiation field reconstruction and uneven surface when generating meshes.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A high-fidelity radiation field reconstruction method based on polarization normal estimation comprises the following steps:

[0006] S1, extracts polarization degree and polarization angle from polarization image sequence, and extracts position code from RGB image sequence;

[0007] S2, inputting the multi-channel tensor consisting of the polarization image sequence, polarization degree, polarization angle and position encoding into the trained attention mechanism image segmentation network to obtain the estimation result of the polarization normal vector; the attention mechanism image segmentation network is an improved network of the image segmentation network Unet++, and the improved network adds an attention gate to the residual connection of the image segmentation network Unet+, and the function expression of the attention gate is:

[0008] ,

[0009] ,

[0010] ,

[0011] in, is the output feature of the attention gate, is the input feature of the attention gate, is the additive attention output, is the output after nonlinear activation and resampling, is the linear transformation matrix, is the ReLU activation function, For input features The weight coefficient of The input gate signal The weight coefficient of is the input gating signal, and is the bias, is the Sigmoid activation function, is the output of the previous layer, For parameters , and A collection of;

[0012] S3, using the motion structure recovery algorithm SfM to extract the camera pose matrix from the RGB image sequence and build a sparse point cloud;

[0013] S4, the estimated result of the polarization normal vector replaces the normal vector of the original DN-Splatter algorithm to obtain an improved DN-Splatter algorithm, and the improved DN-Splatter algorithm is used to achieve high-fidelity radiation field reconstruction based on the established sparse point cloud.

[0014] Optionally, the function expression for extracting the polarization degree and polarization angle from the polarization image sequence in step S1 is:

[0015] ,

[0016] ,

[0017] ,

[0018] ,

[0019] ,

[0020] in, is the degree of polarization, is the polarization angle, degree of polarization It is used to characterize the ratio of the intensity of polarized light in a beam to the total intensity of light, the polarization angle Used to characterize the angle between linear polarization quantities, , and is the Stokes vector, is the polarization intensity in the 0 degree direction, is the polarization intensity in the 45 degree direction, is the polarization intensity in the 90 degree direction.

[0021] Optionally, the function expression for extracting the position code from the RGB image sequence in step S1 is:

[0022] ,

[0023] in, is the position code, and are the focal lengths of the camera in the x and y directions, and are the coordinates of the principal point of the camera in the x and y directions, , and is the coordinate of the real-world point in the camera coordinate system.

[0024] Optionally, before step S2, the method further includes the step of using polarization image sequence samples with polarization normal vector labels to train an attention mechanism image segmentation network, wherein the calculation function expression of the polarization normal vector label is:

[0025] ,

[0026] in, is the polarization normal vector, is a point on the surface of a three-dimensional Gaussian ellipsoid Height, is a point on the surface of a three-dimensional Gaussian ellipsoid The x-axis and y-axis coordinates, is the incident angle, is the azimuth of the incident space, and:

[0027] ,

[0028] ,

[0029] ,

[0030] in, is the refractive index, is the degree of polarization, is the signed inverse tangent function, , and is the Stokes vector, is the polarization intensity in the 0 degree direction, is the polarization intensity in the 45 degree direction, is the polarization intensity in the 90 degree direction, is the polarization intensity in the 135 degree direction.

[0031] Optionally, when the improved DN-Splatter algorithm is used to realize high-fidelity radiation field reconstruction based on the established sparse point cloud, the loss function used is the weighted sum of the photometric error loss and the Gaussian smoothness loss. The function expression of the photometric error loss is:

[0032] ,

[0033] in, is the photometric error loss, is the coefficient, is the target frame, To reconstruct the frame, is the structural similarity measure, and the calculation function expression of the structural similarity measure is:

[0034] ,

[0035] in, express The structural similarity measure of and They are The mean of and They are The variance of for The covariance of and A constant used to avoid division by zero.

[0036] Optionally, the function expression of Gaussian smoothness loss is:

[0037] ,

[0038] in, is the Gaussian smoothness loss, is a pixel, Pixel Depth of neighborhood The gradient of is the transpose operation, Pixel Strength with neighbors gradient.

[0039] In addition, the present invention also provides a high-fidelity radiation field reconstruction system based on polarization normal estimation, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is configured to execute the high-fidelity radiation field reconstruction method based on polarization normal estimation.

[0040] In addition, the present invention also provides a computer-readable storage medium, in which a computer program is stored. The computer program is configured to execute the high-fidelity radiation field reconstruction method based on polarization normal estimation through a processor.

[0041] In addition, the present invention also provides a computer program product, comprising a computer program, wherein the computer program is configured to execute the high-fidelity radiation field reconstruction method based on polarization normal estimation through a processor.

[0042] Compared with the prior art, the present invention has the following advantages: First, in order to improve the accuracy and robustness of 3D reconstruction of scenes and objects, the method of the present invention introduces polarization information into the model reconstruction process. Since the polarization state of light changes when light interacts with the surfaces of different materials, the polarization image can be processed to realize pixel-level normal vector calculation. With the help of deep learning network, the method of the present invention effectively solves the problem of polarization normal vector estimation. The fuzzy problem is solved to ensure the accuracy of multi-object normal vector estimation at the scene level. Secondly, it is very difficult to estimate the normal vector from the sparse points generated by the multi-view geometry of the RGB camera. The method of the present invention uses the normal vector estimated by polarization to provide a priori constraints for the optimization and reconstruction of the radiation field, accelerates the convergence process of the three-dimensional Gaussian ellipsoid optimization, makes the generated mesh surface smoother, and induces the Gaussian sphere to flatten in the normal direction, fits the plane more smoothly, and improves the optimization convergence rate and overall accuracy of the radiation field reconstruction. The present invention has the advantages of strong real-time performance, smooth fitting, and high precision, and can effectively solve the problems of slow convergence of existing radiation field reconstruction and uneven surface when generating meshes. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of the basic flow of the method of the embodiment of the present invention.

[0044] Figure 2The figure is a schematic diagram of the working principle of the method according to the embodiment of the present invention.

[0045] Figure 3 Schematic diagram of the principle of mapping two-dimensional image coordinates to three-dimensional coordinates in an embodiment of the present invention.

[0046] Figure 4 Schematic diagram of the network structure of the attention mechanism image segmentation network in an embodiment of the present invention.

[0047] Figure 5 Schematic diagram of the network structure of the attention gate in an embodiment of the present invention.

[0048] Figure 6 Schematic diagram of the principle of high-fidelity radiation field reconstruction using the improved DN-Splatter algorithm in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] The technical problems to be solved by the high-fidelity radiation field reconstruction method based on polarization normal estimation in the present invention are low accuracy of scene-level polarization normal estimation and uneven surface of radiation field reconstruction, such as Figure 1 and Figure 2 As shown, the high-fidelity radiation field reconstruction method based on polarization normal estimation in this embodiment includes the following steps:

[0051] S1, extracts polarization degree and polarization angle from polarization image sequence, and extracts position code from RGB image sequence;

[0052] S2, input the multi-channel tensor consisting of polarization image sequence, polarization degree, polarization angle and position encoding into the trained attention mechanism image segmentation network to obtain the estimation result of polarization normal vector;

[0053] S3, using the motion structure recovery algorithm SfM to extract the camera pose matrix from the RGB image sequence and build a sparse point cloud;

[0054] S4, the estimated result of the polarization normal vector replaces the normal vector of the original DN-Splatter algorithm to obtain an improved DN-Splatter algorithm, and the improved DN-Splatter algorithm is used to achieve high-fidelity radiation field reconstruction based on the established sparse point cloud.

[0055] The polarization degree and polarization angle information are calculated based on the Stokes vector by inputting polarization images of four polarization directions. The function expression of extracting the polarization degree and polarization angle from the polarization image sequence in step S1 of this embodiment is:

[0056] ,

[0057] ,

[0058] ,

[0059] ,

[0060] ,

[0061] in, is the degree of polarization, is the polarization angle, degree of polarization It is used to characterize the ratio of the intensity of polarized light in a beam to the total intensity of light, the polarization angle Used to characterize the angle between linear polarization quantities, , and is the Stokes vector, is the polarization intensity in the 0 degree direction, is the polarization intensity in the 45 degree direction, is the polarization intensity in the 90 degree direction, is the pixel coordinate

[0062] Since the position information of the observation direction will affect the measured polarization information, the two-dimensional image coordinates are mapped to three dimensions through the camera intrinsic parameters and normalized to form a position code. Figure 3 Schematic diagram of the principle of mapping two-dimensional image coordinates to three-dimensional coordinates in this embodiment, where O is the origin, is the two-dimensional image coordinate, is a three-dimensional coordinate. In step S1 of this embodiment, the function expression for extracting the position code from the RGB image sequence is:

[0063] ,

[0064] in, is the position code, and are the focal lengths of the camera in the x and y directions, and are the coordinates of the principal point of the camera in the x and y directions, , and is the coordinate of the real-world point in the camera coordinate system.

[0065] The multi-channel tensor consisting of the original polarization image, polarization degree, polarization angle and position encoding is input into the trained attention mechanism image segmentation network to obtain the estimated result of the polarization normal vector. Considering the difficulty of extracting normal vectors of multiple objects in complex scenes, see Figure 4 In this embodiment, an attention gate is added to the residual connection of the image segmentation network Unet++ with dense skip connections ( Figure 4 The triangle in the figure is the added attention gate. The rest of the structure is the same as Unet++. The X in each circular frame is i,j Representing features), it can connect the feature information of the upper and lower layers, enhance the generalization of the network and the interactivity between features, allow the network to learn features of different depths by itself, and at the same time control the number of parameters to a certain extent to prevent overfitting in the case of small sample training. The input of the attention gate is the upsampled features of the extended node and the corresponding features from the current encoder. The upsampled features are used as gating signals to suppress irrelevant areas in the task. The newly designed network effectively reduces the angular error of normal vector estimation, improves the scene generalization of normal vector estimation, and provides accurate prior constraints for 3D reconstruction tasks. Figure 5 This is a schematic diagram of the network structure of the attention gate in this embodiment. Assume that the current encoder feature of the input is , the sampling feature on the expanded node is , the function expression of the attention gate is:

[0066] ,

[0067] ,

[0068] ,

[0069] in, For attention The output characteristics of the gate, is the input feature of the attention gate, is the additive attention output, is the output after nonlinear activation and resampling, is the linear transformation matrix, is the ReLU activation function, For input features The weight coefficient of The input gate signal The weight coefficient of is the input gating signal, and is the bias, is the Sigmoid activation function, is the output of the previous layer, For parameters , and Resampling refers to the output and weight matrix of the previous node. Multiply elements by element to enhance features.

[0070] Before step S2, this embodiment also includes a step of using polarization image sequence samples with polarization normal vector labels to train an attention mechanism image segmentation network, wherein the calculation function expression of the polarization normal vector label is:

[0071] ,

[0072] in, is the polarization normal vector, is a point on the surface of a three-dimensional Gaussian ellipsoid Height, is a point on the surface of a three-dimensional Gaussian ellipsoid The x-axis and y-axis coordinates, is the incident angle, is the azimuth of the incident space, and:

[0073] ,

[0074] ,

[0075] ,

[0076] in, is the refractive index, is the degree of polarization, is the signed inverse tangent function, , and is the Stokes vector, is the polarization intensity in the 0 degree direction, is the polarization intensity in the 45 degree direction, is the polarization intensity in the 90 degree direction, is the polarization intensity in the 135 degree direction.

[0077] See also Figure 2In step S3 of this embodiment, the RGB image sequence is extracted from the camera pose matrix using the motion structure recovery algorithm SfM and a sparse point cloud is established. The motion structure recovery algorithm SfM is an existing algorithm (see the literature: Snavely, N., Seitz, SM, & Szeliski, R. (2006). Photo Tourism: Exploring PhotoCollections In 3D. ACM Transactions on Graphics, 25(3), 835-846.). The algorithm includes two core links: feature extraction and matching and incremental 3D reconstruction. A set of images is input, and feature points are extracted from each input image using the feature extraction and matching algorithm, and descriptor matching is performed. The RANSAC algorithm is used to eliminate mismatches and estimate the geometric model (basic matrix or essential matrix) between cameras; the incremental 3D reconstruction is to preliminarily restore the camera pose and select image pairs for initial registration. A new view is added, and a sparse point cloud is established by triangulation of matching feature points for building an incremental map, and a redundancy and topology verification mechanism is used to improve the reconstruction accuracy and robustness.

[0078] In step S4 of this embodiment, the estimated result of the polarization normal vector replaces the normal vector of the original DN-Splatter algorithm (see reference: Bhat, SF, Birkl, R., Wofk, D., Wonka, P.,&Müller, M. (2023). DN-Splatter: Depth and Normal Priors for Gaussian Splatting and Meshing. arXiv.arXiv:2403.17822) to obtain an improved DN-Splatter algorithm, and the improved DN-Splatter algorithm is used to achieve high-fidelity radiation field reconstruction based on the established sparse point cloud. Figure 6 This is a schematic diagram of the principle of high-fidelity radiation field reconstruction of the improved DN-Splatter algorithm in this embodiment. In this embodiment, the normal vector is replaced by the estimated result of the polarization normal vector, and the corresponding Gaussian ellipsoid is constrained to expand in the normal vector direction according to the polarization normal vector of the pixel point, so as to fit the real surface in a flattened form as much as possible. The Gaussian ellipsoid optimization variables mainly include information such as size, color, and posture in the three-axis directions. The normal vector constraint regularizes the Gaussian position of the edge perception loss, fits the surface Gaussian local smoothing and direction correction during the optimization process.

[0079] As an optional implementation, when the improved DN-Splatter algorithm is used in this embodiment to realize high-fidelity radiation field reconstruction according to the established sparse point cloud, the loss function adopted is the weighted sum of the photometric error loss and the Gaussian smoothness loss. The function expression of the photometric error loss is:

[0080] ,

[0081] in, is the photometric error loss, is the coefficient, is the target frame, To reconstruct the frame, is the structural similarity measure, and the calculation function expression of the structural similarity measure is:

[0082] ,

[0083] in, express The structural similarity measure of and They are The mean of and They are The variance of for The covariance of and A constant used to avoid division by zero.

[0084] In this embodiment, the function expression of Gaussian smoothness loss is:

[0085] ,

[0086] in, is the Gaussian smoothness loss, is a pixel, Pixel Depth of neighborhood The gradient of is the transpose operation, Pixel Strength with neighbors gradient.

[0087] In summary, the high-fidelity radiation field reconstruction method based on polarization normal estimation in this embodiment has the following characteristics: 1) This embodiment method first proposes a new method for using the normal vector of polarization estimation to guide the reconstruction of three-dimensional radiation fields, which accelerates the convergence rate of radiation field optimization and greatly improves the fitting accuracy of the three-dimensional reconstructed surface. 2) This embodiment method improves the existing scene-level polarization normal vector estimation network in the jump connection of Unet++ to enhance the learning of the target area, so that the attention connection between adjacent convolutional layers better handles the relationship between local normal vector estimation and overall scene segmentation. 3) This embodiment method can estimate an accurate normal vector map only through a single-view polarization image, eliminating the step of using the normal direction of the estimated initial SfM point cloud to initialize the Gaussian direction for monocular normal estimation. 4) This embodiment method can make the mesh further generated by three-dimensional radiation field reconstruction smoother, which can be better suitable for downstream robot navigation tasks.

[0088] In addition, this embodiment also provides a high-fidelity radiation field reconstruction system based on polarization normal estimation, including a microprocessor and a memory connected to each other, and the microprocessor is configured to execute the high-fidelity radiation field reconstruction method based on polarization normal estimation.

[0089] In addition, this embodiment further provides a computer-readable storage medium, in which a computer program is stored. The computer program is configured to execute the high-fidelity radiation field reconstruction method based on polarization normal estimation through a processor.

[0090] In addition, this embodiment also provides a computer program product, including a computer program, where the computer program is configured to execute the high-fidelity radiation field reconstruction method based on polarization normal estimation through a processor.

[0091] Those skilled in the art should understand that the technical solutions provided by the embodiments of the present application may be in the form of methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the process Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0092] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A high-fidelity radiation field reconstruction method based on polarization normal estimation, characterized in that: The steps include: S1, extracts polarization degree and polarization angle from polarization image sequence, and extracts position code from RGB image sequence; S2, inputting the multi-channel tensor consisting of the polarization image sequence, polarization degree, polarization angle and position encoding into the trained attention mechanism image segmentation network to obtain the estimation result of the polarization normal vector; the attention mechanism image segmentation network is an improved network of the image segmentation network Unet++, and the improved network adds an attention gate to the residual connection of the image segmentation network Unet+, and the function expression of the attention gate is: ， ， ， in, is the output feature of the attention gate, is the input feature of the attention gate, is the additive attention output, is the output after nonlinear activation and resampling, is the linear transformation matrix, is the ReLU activation function, For input features The weight coefficient of The input gate signal The weight coefficient of is the input gating signal, and is the bias, is the Sigmoid activation function, is the output of the previous layer, For parameters , and A collection of; S3, using the motion structure recovery algorithm SfM to extract the camera pose matrix from the RGB image sequence and build a sparse point cloud; S4, the estimated result of the polarization normal vector replaces the normal vector of the original DN-Splatter algorithm to obtain an improved DN-Splatter algorithm, and the improved DN-Splatter algorithm is used to achieve high-fidelity radiation field reconstruction based on the established sparse point cloud.

2. The high-fidelity radiation field reconstruction method based on polarization normal estimation according to claim 1, characterized in that: The functional expression for extracting the polarization degree and polarization angle from the polarization image sequence in step S1 is: ， ， ， ， ， in, is the degree of polarization, is the polarization angle, degree of polarization It is used to characterize the ratio of the intensity of polarized light in a beam to the total intensity of light, the polarization angle Used to characterize the angle between linear polarization quantities, , and is the Stokes vector, is the polarization intensity in the 0 degree direction, is the polarization intensity in the 45 degree direction, is the polarization intensity in the 90 degree direction.

3. The high-fidelity radiation field reconstruction method based on polarization normal estimation according to claim 1, characterized in that: The function expression for extracting the position code from the RGB image sequence in step S1 is: ， in, is the position code, and are the focal lengths of the camera in the x and y directions, and are the coordinates of the principal point of the camera in the x and y directions, , and is the coordinate of the real-world point in the camera coordinate system.

4. The high-fidelity radiation field reconstruction method based on polarization normal estimation according to claim 1, characterized in that: Before step S2, the method also includes the step of using polarization image sequence samples with polarization normal vector labels to train the attention mechanism image segmentation network, wherein the calculation function expression of the polarization normal vector label is: ， in, is the polarization normal vector, is a point on the surface of a three-dimensional Gaussian ellipsoid Height, is a point on the surface of a three-dimensional Gaussian ellipsoid The x-axis and y-axis coordinates, is the incident angle, is the azimuth of the incident space, and: ， ， ， in, is the refractive index, is the degree of polarization, is the signed inverse tangent function, , and is the Stokes vector, is the polarization intensity in the 0 degree direction, is the polarization intensity in the 45 degree direction, is the polarization intensity in the 90 degree direction, is the polarization intensity in the 135 degree direction.

5. The high-fidelity radiation field reconstruction method based on polarization normal estimation according to claim 4, characterized in that: When the improved DN-Splatter algorithm is used to achieve high-fidelity radiation field reconstruction based on the established sparse point cloud, the loss function used is the weighted sum of the photometric error loss and the Gaussian smoothness loss. The function expression of the photometric error loss is: ， in, is the photometric error loss, is the coefficient, is the target frame, To reconstruct the frame, is the structural similarity measure, and the calculation function expression of the structural similarity measure is: ， in, express The structural similarity measure of and They are The mean of and They are The variance of for The covariance of and A constant used to avoid division by zero.

6. The high-fidelity radiation field reconstruction method based on polarization normal estimation according to claim 5, characterized in that: The function expression of Gaussian smoothness loss is: ， in, is the Gaussian smoothness loss, is a pixel, Pixel Depth of neighborhood The gradient of is the transpose operation, Pixel Strength with neighbors gradient.

7. A high-fidelity radiation field reconstruction system based on polarization normal estimation, comprising a microprocessor and a memory connected to each other, characterized in that: The microprocessor is configured to execute the high-fidelity radiation field reconstruction method based on polarization normal estimation as recited in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored therein, characterized in that: The computer program is configured to execute the high-fidelity radiation field reconstruction method based on polarization normal estimation as claimed in any one of claims 1 to 6 through a processor.

9. A computer program product, comprising a computer program, characterized in that The computer program is configured to execute the high-fidelity radiation field reconstruction method based on polarization normal estimation as claimed in any one of claims 1 to 6 through a processor.

Images