High-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization

Abstract

The invention relates to a high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization, and belongs to the field of computer vision and graphic images. The method comprises the following steps: firstly, constructing a fuzzy sensing module, accurately identifying a fuzzy type and an area of an input multi-view fuzzy image, and generating a corresponding fuzzy mask; performing three-dimensional Gaussian representation based on the input multi-view blurred image, estimating internal and external parameters of a camera and sparse point cloud data of a scene, and obtaining initial three-dimensional Gaussian scene representation; and then performing joint optimization on a Gaussian ellipsoid in the initial three-dimensional Gaussian scene representation in combination with an input multi-view blurred image and a blurred mask to generate a clear three-dimensional scene representation. And finally, performing three-dimensional Gaussian reconstruction rendering by using the optimized three-dimensional Gaussian scene representation, and outputting a high-precision three-dimensional Gaussian reconstruction result and a clear rendered image. The method can effectively improve the three-dimensional reconstruction precision and rendering quality under the fuzzy input condition, has high adaptability and stability, and can be applied to the fields of image deblurring, virtual reality and the like.

Description

Technical Field

[0001] The present invention relates to a high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization, and belongs to the field of computer vision and graphics. Background Art

[0002] Fuzzy Multi-View Figure 3 3D reconstruction aims to reconstruct clear and high-quality 3D scenes from multi-view blurred image sequences to improve the robustness of 3D reconstruction methods. It is one of the important problems in the field of computer vision and graphics, and has broad application prospects in digital applications such as robot navigation, AR / VR / MR, e-commerce and 3D games. Most of the known methods focus on specific types such as defocus blur or camera motion blur. Figure 3 This type of method can achieve good results in the 3D reconstruction task of a specific type of blur image, but because the model is only for a single type of blur, when other types of blur are included, it cannot be processed, resulting in incorrect fitting to the 3D model and low reconstruction quality. In addition, the blur kernel-based method generalizes the 2D blur kernel to 3D space, realizing fuzzy multi-viewing. Figure 3 dimensional reconstruction. For example, Wu(<ACMInternational Conference on Multimedia> ,2022) proposed DoF-NeRF, which achieves 3D reconstruction based on defocus blur by learning aperture size and focal length parameters and using optically consistent scattering to model scene depth.<International Conference on Computer Vision> ,2023) modeled the motion trajectory of a 6-DOF camera based on Bezier curves, and accumulated light along the estimated camera trajectory to obtain blurred RGB colors, thereby optimizing the 3D reconstruction results of the camera motion blurred image.<European Conference on Computer Vision> ,2024) introduced the 2D degradation model Blur Proposal Network (BPN), modeled the blur by the blur kernel of each pixel estimated by BPN, and used the quality assessment mask to mark the blur occurrence area, optimized the blur kernel from coarse to fine, and realized 3D reconstruction. The above-mentioned known methods all deal with multi-view of a single blur type. Figure 3 However, in real blurry scenes, different blur types have different spatial distributions and physical causes. The use of a single blur imaging mechanism or a uniform blur kernel in the known methods not only increases the learning difficulty but also causes geometric distortion and reduces the quality of 3D reconstruction.

[0003] Fuzzy multi-viewing designed for a single blur type Figure 3The 3D reconstruction strategy implicitly assumes fixed degradation, which makes such methods have insufficient generalization ability when facing multi-view image sequence input containing camera motion blur and defocus blur. The known methods cannot effectively model the complex degradation mechanism across blur types, resulting in incomplete blur removal, leading to blur and artifacts in the reconstruction results; the lack of blur consistency between multi-view images with camera motion blur and defocus blur also affects the registration and fusion accuracy between multiple views in 3D reconstruction. The non-uniformity of the blur kernel in the existing known methods causes the same scene point to present different blur responses in different images, violating the photometric consistency assumption, thereby exacerbating the instability of disparity estimation. Therefore, the present invention first constructs a blur perception module to accurately identify the blur type and area of ​​the input multi-view blurred image and generate a corresponding blur mask; then, a 3D Gaussian representation is performed based on the input multi-view blurred image to obtain an initial 3D Gaussian scene representation; then, the Gaussian ellipsoid in the initial 3D Gaussian scene representation is jointly optimized by combining the input multi-view blurred image and the blur mask to generate a clear 3D scene representation, thereby improving the 3D reconstruction accuracy and rendering quality under blurry input conditions. Summary of the Invention

[0004] The present invention provides a high-precision 3D Gaussian reconstruction method based on blur perception and joint optimization, which is used to solve the problems of camera global pose estimation errors caused by the coexistence of camera motion blur and defocused blurred images in the input multi-view image sequence of 3D reconstruction, and the presence of geometric distortion and artifacts in the reconstruction results, thereby generating high-quality 3D reconstruction results and clear rendered images.

[0005] The technical solution of the present invention is: a high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization, comprising the following steps:

[0006] Step 1: Build a blur perception module to identify the blur type and area of ​​the input multi-view blur image and generate the corresponding blur mask;

[0007] Step 2: Perform a 3D Gaussian representation based on the input multi-view blurred image, estimate the camera intrinsic and extrinsic parameters, and the sparse point cloud data of the scene to obtain an initial 3D Gaussian scene representation;

[0008] Step 3: Combine the input multi-view blurred image and the blur mask to jointly optimize the Gaussian ellipsoid in the initial 3D Gaussian scene representation to generate an optimized 3D Gaussian scene representation;

[0009] Step 4: Use the optimized 3D Gaussian scene representation to perform 3D Gaussian reconstruction rendering, and output high-precision 3D Gaussian reconstruction results and clear rendered images.

[0010] As a further solution of the present invention, the specific process of Step 1 is as follows:

[0011] Step 1.1 Fuzzy type perception;

[0012] First, define the fuzzy type perception network C btp , for the multi-view blurred image dataset containing camera motion blur and defocus blur {I blur}Through C btp Perform forward propagation to obtain the image blur type label L, where L∈{motion,defocus}, motion and defocus represent camera motion blur and defocus blur, respectively;

[0013] Then, for the image with label L as motion Generate and All-1 binary mask M with consistent resolution region ;

[0014] Step 1.2 Fuzzy area perception;

[0015] First, for the image labeled L as defocus Perform feature extraction to obtain high-level semantic features F of the defocused image;

[0016] Secondly, the transposed convolution operation is applied to F to increase the feature size layer by layer. After 6 layers of decoding, the original resolution of the image is restored to generate a preliminary blur mask M. dec ;

[0017] Then, we construct the deep feature aggregation module M DFA =ReLU(h(Concat(F deep1 ,F deep2 ,F deep3 ))) and shallow feature aggregation module M SFA =ReLU(h(Concat(F shallow1 ,F shallow2 ,F shallow3 ))), aggregate deep and shallow features to obtain the fuzzy global structure information mask M DFA and blurred local texture information mask M SFA , where h(·) is a 1×1 convolution operation, and Concat(·) represents a feature fusion operation;

[0018] Finally, the fusion feature mask M dec ,M DFA ,M SFA Generate defocused blur area mask M region , and its fusion process is expressed as: M region =σ(h(Concat(M dec ,M DFA,M SFA ))), where σ(·) is the Sigmoid activation function, h(·) is the 1×1 convolution operation, and Concat(·) represents the feature fusion operation.

[0019] As a further solution of the present invention, the specific process of Step 3 is as follows:

[0020] Step 3.1 Blur view rendering;

[0021] First, based on the camera internal and external parameters obtained in Step 2, estimate the starting position T of the camera motion trajectory s and the ending pose T e ;

[0022] Then, linear interpolation is used to calculate T s and T e Generate n virtual interpolated camera poses T in SE(3) space i ,Right now

[0023] Finally, based on T i From the initial 3D Gaussian scene representation, we can obtain the image I corresponding to each virtual camera pose by projection and differentiable rasterization rendering. ri , get the fused rendered image

[0024] Step 3.2 Combined opacity and UGR optimization;

[0025] First, for the input image I l and rendered image I rl Use the mask of the lth view Perform mask operations and obtain and is the input image I of the lth view l The result of the mask operation, The rendering method of step 3.1 is used to render the 3D Gaussian model to obtain the rendered image I of the lth view. rl The result of the mask operation; and After calculating the loss, the camera pose and 3D Gaussian parameters are optimized through gradient backpropagation;

[0026] Then, define the unoptimized Gaussian ratio Calculate the optimization of all Gaussian ellipsoids within the optimization cycle, where M is the optimization cycle, O j ∈{0,1} indicates whether the Gaussian ellipsoid is optimized in the jth iteration within the optimization cycle; the proposed distribution of state jumps is constructed by combining the opacity of the Gaussian ellipsoid with the UGR Perform joint optimization, where gi represents the Gaussian ellipsoid with index i, σ i Indicates the transparency of the Gaussian ellipsoid with index i, ω is the balance factor, and Z is the normalization term;

[0027] Finally, the opacity σ i The Gaussian ellipsoid below the set death threshold is judged as "death Gaussian". i ) performs polynomial sampling, selects the target Gaussian from the current active Gaussian set, relocates the "dead Gaussian" to the spatial neighborhood of the selected target Gaussian, and generates a clear three-dimensional Gaussian scene representation G.

[0028] The present invention also provides a high-precision three-dimensional Gaussian reconstruction system based on fuzzy perception and joint optimization, the system comprising: a module for executing the high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization.

[0029] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, the high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization is implemented.

[0030] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization is implemented.

[0031] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization.

[0032] The beneficial effects of the present invention are:

[0033] 1. Conventional methods primarily reconstruct single blurred images based on blur formation or blur kernel mechanisms. These methods are unable to simultaneously process input multi-view image sequences containing camera motion blur and defocus blur, leading to artifacts and blur in the reconstruction and rendering results. This invention, by constructing a blur perception module, accurately identifies the blur type and region of the input multi-view blurred image and generates a corresponding pixel-level blur mask to guide 3D reconstruction, improving the generalization capability for real-world blurred scenes.

[0034] 2. Known methods are usually based on 3D Gaussian splatting for fuzzy multi-viewing Figure 33D Gaussian reconstruction cannot effectively use blur perception results to guide 3D Gaussian reconstruction. This paper uses a blur-perceived 3D Gaussian splattering framework to handle camera motion blur and defocus blur, and uses a blur mask to guide model gradient optimization, ensuring accurate modeling and avoiding degradation.

[0035] 3. Most known methods use heuristic adaptive density control strategies to optimize 3D Gaussians, which are difficult to adapt to different scenarios and are prone to redundant Gaussian points during training, resulting in high computational costs. This method combines the input multi-view blurred image and blur mask, and performs Gaussian relocalization by combining the opacity of the Gaussian ellipsoid with the UGR to construct a state jump proposal distribution, guiding the adaptive adjustment of the model structure. This enhances the modeling and optimization capabilities of blurred regions and improves the reconstruction quality of weakly textured areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the present invention;

[0037] Figure 2 This is the flowchart of fuzzy area perception;

[0038] Figure 3 Optimizing the flow chart for blur-aware 3D Gaussian splattering;

[0039] Figure 4 Optimize the schematic diagram for combined opacity and UGR;

[0040] Figure 5 Fuzzy multi-viewing Figure 3 Dimensional reconstruction rendering; DETAILED DESCRIPTION

[0041] Example 1: Figure 1-Figure 5 As shown in FIG, a high-precision 3D Gaussian reconstruction method based on fuzzy perception and joint optimization includes:

[0042] Step 1: Build a blur perception module to accurately identify the blur type and area of ​​the input multi-view blur image and generate the corresponding blur mask;

[0043] Step 2: Perform a 3D Gaussian representation based on the input multi-view blurred image, estimate the camera intrinsic and extrinsic parameters, and the sparse point cloud data of the scene to obtain an initial 3D Gaussian scene representation;

[0044] Step 3: Combine the input multi-view blurred image and the blur mask to jointly optimize the Gaussian ellipsoid in the initial 3D Gaussian scene representation to generate a clear 3D scene representation;

[0045] Step 4: Use the optimized 3D Gaussian scene representation to perform 3D Gaussian reconstruction rendering, and output high-precision 3D Gaussian reconstruction results and clear rendered images.

[0046] The specific process of Step 1 is as follows:

[0047] Step 1.1 Fuzzy type perception

[0048] First, define the fuzzy type perception network C btp , for the multi-view blurred image dataset containing camera motion blur and defocus blur {I blur}Through C btp Perform forward propagation to obtain the image blur type label L, where L∈{motion,defocus}, motion and defocus represent camera motion blur and defocus blur, respectively;

[0049] Then, for the image labeled L as motion Generate and All-1 binary mask M with consistent resolution region .

[0050] Step 1.2 Fuzzy Area Perception

[0051] First, for the image labeled L as defocus Perform feature extraction to obtain high-level semantic features F of the defocused image;

[0052] Secondly, the transposed convolution operation is applied to F to increase the feature size layer by layer. After 6 layers of decoding, the original resolution of the image is restored to generate a preliminary blur mask M. dec ;

[0053] Then, we construct the deep feature aggregation module M DFA =ReLU(h(Concat(F deep1 ,F deep2 ,F deep3 ))) and shallow feature aggregation module M SFA =ReLU(h(Concat(F shallow1 ,F shallow2 ,F shallow3 ))), aggregate deep and shallow features to obtain the fuzzy global structure information mask M DFA and blurred local texture information mask M SFA , where h(·) is a 1×1 convolution operation and Concatσ(·) represents a feature fusion operation;

[0054] Finally, the fusion feature mask M dec ,M DFA ,M SFA Generate defocused blur area mask M region , and its fusion process is expressed as: M region=σ(h(Concat(M dec ,M DFA ,M SFA ))), where σ(·) is the Sigmoid activation function, hσ(·) is the 1×1 convolution operation, and Concat(·) represents the feature fusion operation.

[0055] The specific process of Step 3 is as follows:

[0056] Step 3.1 Blur View Rendering

[0057] First, based on the camera internal and external parameters obtained in Step 2, estimate the starting position T of the camera motion trajectory s and the ending pose T e ;

[0058] Then, linear interpolation is used to calculate T s and T e Generate n virtual interpolated camera poses T in SE(3) space i ,Right now

[0059] Finally, based on T i From the initial 3D Gaussian scene representation, we can obtain the image I corresponding to each virtual camera pose by projection and differentiable rasterization rendering. ri , get the fused rendered image

[0060] Step 3.2 Combined opacity and UGR optimization

[0061] First, for the input image I l and rendered image I rl Use the mask of the lth view Perform mask operations and obtain and is the input image I of the lth view l The result of the mask operation, The rendering method of step 3.1 is used to render the 3D Gaussian model to obtain the rendered image I of the lth view. rl The result of the mask operation; and After calculating the loss, the camera pose and 3D Gaussian parameters are optimized through gradient backpropagation;

[0062] Then, define the unoptimized Gaussian ratio Calculate the optimization of all Gaussian ellipsoids within the optimization cycle, where M is the optimization cycle, O j∈{0,1} indicates whether the Gaussian ellipsoid is optimized in the jth iteration within the optimization cycle. The proposed distribution of state jumps is constructed by combining the opacity of the Gaussian ellipsoid with the UGR. Perform joint optimization, where g i represents the Gaussian ellipsoid with index i, σ i Indicates the transparency of the Gaussian ellipsoid with index i, ω is the balance factor, and Z is the normalization term;

[0063] Finally, the opacity σ i The Gaussian ellipsoid below the set death threshold is judged as "death Gaussian". i ) performs polynomial sampling, selects the target Gaussian from the current active Gaussian set, relocates the “dead Gaussian” to the spatial neighborhood of the selected target Gaussian, and generates a clear three-dimensional scene representation G.

[0064] The present invention also provides a high-precision three-dimensional Gaussian reconstruction system based on fuzzy perception and joint optimization, the system comprising: a module for executing the high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization.

[0065] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, the high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization is implemented.

[0066] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization is implemented.

[0067] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization.

[0068] Example 2: Figure 1 As shown, the specific steps of this method are as follows:

[0069] Step 1.1, perform fuzzy type perception. First, define the fuzzy type perception network C based on the YOLO11 classification network. btp , for the input multi-view blurred image dataset containing multi-view images with camera motion blur and defocus blur By C btp Perform forward propagation to obtain the image blur type label L, where L∈{motion,defocus}, motion and defocus represent camera motion blur and defocus blur, respectively;

[0070] Then, for the image labeled L as defocused Use OpenCV to generate All-1 binary mask with consistent resolution

[0071] Step 1.2 Figure 2 As shown, the blurred area perception is performed. First, ResNet50 is used as the feature extraction network to extract the image with label L as camera motion blur. First, normalization is performed, and then the feature F is obtained through convolution layer, batch normalization, ReLU activation and pooling. (0) , then F (0) The feature extraction F is carried out in four stages in sequence. l =H l (F l-1 θ l ), and obtain the high-level semantic features of the defocused image Where, l∈{1,2,3,4},F l represents the feature map of layer l, H l Represents the convolution-normalization-activation function operation of layer l, θ l are network parameters, H and W are input image sizes;

[0072] Secondly, for F( 4 ) uses the transposed convolution operation to increase the feature size layer by layer, and restores the original resolution of the image through 6 layers of decoding to generate a preliminary blur mask M dec ;

[0073] Then, we construct the deep feature aggregation module M DFA =ReLU(h(Concat(F deep1 ,F deep2 ,F deep3 ))) and shallow feature aggregation module M SFA =ReLU(h(Concat(F shallow1 ,F shallow2 ,F shallow3 ))), aggregate deep and shallow features to obtain the fuzzy global structure information mask M DFA and blurred local texture information mask M SFA , where h(·) is a 1×1 convolution operation, and Concat(·) represents a feature fusion operation;

[0074] Finally, the fusion feature mask M dec ,M DFA ,M SFA Generate defocused blur area mask The fusion process is expressed as: Mregion =σ(h(Concat(M dec ,M DFA ,M SFA ))), where σ(·) is the Sigmoid activation function, h(·) is the 1×1 convolution operation, and Concat(·) represents the feature fusion operation.

[0075] Step 2: Use COLMAP to estimate the camera intrinsic and extrinsic parameters based on the input multi-view blurred image, and perform a 3D Gaussian representation on the sparse point cloud data of the scene to obtain an initial 3D Gaussian scene representation;

[0076] Step 3.1 performs fuzzy view rendering. First, based on the camera internal and external parameters obtained in Step 2, the starting position T of the camera motion trajectory is estimated. s and the ending pose T e ;

[0077] Then, linear interpolation is used to calculate T s and T e Generate n virtual interpolated camera poses T in SE(3) space i ,Right now

[0078] Finally, based on T i From the initial 3D Gaussian scene representation, we can obtain the image I corresponding to each virtual camera pose by projection and differentiable rasterization rendering. ri , get the fused rendered image

[0079] Step 3.2 Combined opacity and UGR optimization process Figure 3 As shown. First, for the input image I l and rendered image I rl Use the mask of the lth view Perform mask operations and obtain and is the input image I of the lth view l The result of the mask operation, The rendering method of step 3.1 is used to render the 3D Gaussian model to obtain the rendered image I of the lth view. rl The result of the mask operation; and After calculating the loss, the camera pose and 3D Gaussian parameters are optimized through gradient backpropagation;

[0080] Then, based on the Bayesian inference idea, the three-dimensional Gaussian scene representation is regarded as a set of samples sampled from the posterior distribution. By constructing a Markov Chain Monte Carlo (MCMC) sampling process and using the noisy gradient update (SGLD) to simulate the sampling behavior of the parameters, the global exploration and adjustment of the distribution are carried out to achieve the unified goals of controlling the number of Gaussians, optimizing the spatial distribution, and improving the convergence stability. The specific process is as follows: By defining the unoptimized Gaussian ratio Calculate the optimization of all Gaussian ellipsoids within the optimization cycle, where M is the optimization cycle, O j ∈{0,1} indicates whether the Gaussian ellipsoid is optimized in the jth iteration within the optimization cycle. The proposed distribution of state jumps is constructed by combining the opacity of the Gaussian ellipsoid with the UGR. Perform joint optimization, where g i represents the Gaussian ellipsoid with index i, σ i Indicates the transparency of the Gaussian ellipsoid with index i, ω is the balance factor, and Z is the normalization term;

[0081] Finally, if Figure 4 As shown, the opacity σ i The Gaussian ellipsoid below the set death threshold is judged as "death Gaussian". i ) performs polynomial sampling, selects the target Gaussian from the current active Gaussian set, relocates the “dead Gaussian” to the spatial neighborhood of the selected target Gaussian, and generates a clear three-dimensional scene representation G.

[0082] Step 4: Based on the optimized 3D Gaussian representation G, project the center point coordinates, scaling factor (spatial covariance), orientation (quaternion), opacity, and color information of the 3D Gaussian onto the image plane, calculate the 2D Gaussian distribution, and use a forward fusion strategy to perform weighted accumulation of each Gaussian based on the principles of differentiable rasterization and alpha blending. The pixel color is determined by the following formula: Among them, α i is the opacity of the i-th Gaussian ellipsoid, c i is its color value. Finally, we get high-precision 3D Gaussian reconstruction results and clear rendering images. The multi-view blurred image reconstruction rendering results are shown in Figure 5 , including 28 blurred view images caused by camera motion blur and defocus blur, four input view angles, and four new view angles. The clear images rendered from the original and new view angles demonstrate the high accuracy of the reconstruction results, with good restoration of scene details and almost no artifacts.

[0083] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A high-precision 3D Gaussian reconstruction method based on fuzzy perception and joint optimization, characterized by: The steps include: Step 1: Build a blur perception module to identify the blur type and area of ​​the input multi-view blur image and generate the corresponding blur mask; Step 2: Perform a 3D Gaussian representation based on the input multi-view blurred image, estimate the camera intrinsic and extrinsic parameters, and the sparse point cloud data of the scene to obtain an initial 3D Gaussian scene representation; Step 3: Combine the input multi-view blurred image and the blur mask to jointly optimize the Gaussian ellipsoid in the initial 3D Gaussian scene representation to generate an optimized 3D Gaussian scene representation; Step 4: Use the optimized 3D Gaussian scene representation to perform 3D Gaussian reconstruction rendering, and output high-precision 3D Gaussian reconstruction results and clear rendered images.

2. The high-precision 3D Gaussian reconstruction method based on fuzzy perception and joint optimization according to claim 1, characterized in that: The specific process of Step 1 is as follows: Step 1.1 Fuzzy type perception; First, define the fuzzy type perception network C btp , for the multi-view blurred image dataset containing camera motion blur and defocus blur {I blur }Through C btp Perform forward propagation to obtain the image blur type label L, where L∈{motion,defocus}, motion and defocus represent camera motion blur and defocus blur, respectively; Then, for the image with label L as motion Generate and All-1 binary mask M with consistent resolution region ; Step 1.2 Fuzzy area perception; First, for the image labeled L as defocus Perform feature extraction to obtain high-level semantic features F of the defocused image; Secondly, the transposed convolution operation is applied to F to increase the feature size layer by layer. After 6 layers of decoding, the original resolution of the image is restored to generate a preliminary blur mask M. dec ; Then, we construct the deep feature aggregation module M DFA =ReLU(h(Concat(F deep1 ,F deep2 ,F deep3 ))) and shallow feature aggregation module M SFA =ReLU(h(Concat(F shallow1 ,F shallow2 ,F shallow3 ))), aggregate deep and shallow features to obtain the fuzzy global structure information mask M DFA and blurred local texture information mask M SFA , where h(·) is a 1×1 convolution operation, and Concat(·) represents a feature fusion operation; Finally, the fusion feature mask M dec ,M DFA ,M SFA Generate defocused blur area mask M region , and its fusion process is expressed as: M region =σ(h(Concat(M dec ,M DFA ,M SFA ))), where σ(·) is the Sigmoid activation function, h(·) is the 1×1 convolution operation, and Concat(·) represents the feature fusion operation.

3. The high-precision 3D Gaussian reconstruction method based on fuzzy perception and joint optimization according to claim 1, characterized in that: The specific process of Step 3 is as follows: Step 3.1 Blur view rendering; First, based on the camera internal and external parameters obtained in Step 2, estimate the starting position T of the camera motion trajectory s and the ending pose T e ; Then, linear interpolation is used to calculate T s and T e Generate n virtual interpolated camera poses T in SE(3) space i ,Right now Finally, based on T i From the initial 3D Gaussian scene representation, we can obtain the image I corresponding to each virtual camera pose by projection and differentiable rasterization rendering. ri , get the fused rendered image Step 3.2 Combined opacity and UGR optimization; First, for the input image I l and rendered image I rl Use the mask of the lth view Perform mask operations and obtain and is the input image I of the lth view l The result of the mask operation, The rendering method of step 3.1 is used to render the 3D Gaussian model to obtain the rendered image I of the lth view. rl The result of the mask operation; and After calculating the loss, the camera pose and 3D Gaussian parameters are optimized through gradient backpropagation; Then, define the unoptimized Gaussian ratio Calculate the optimization of all Gaussian ellipsoids within the optimization cycle, where is the optimization cycle, O j ∈{0,1} indicates whether the Gaussian ellipsoid is optimized in the jth iteration within the optimization cycle; the proposed distribution of state jumps is constructed by combining the opacity of the Gaussian ellipsoid with the UGR Perform joint optimization, where g i represents the Gaussian ellipsoid with index i, σ i Indicates the transparency of the Gaussian ellipsoid with index i, ω is the balance factor, and Z is the normalization term; Finally, the opacity σ i The Gaussian ellipsoid below the set death threshold is judged as "death Gaussian". i ) performs polynomial sampling, selects the target Gaussian from the current active Gaussian set, relocates the "dead Gaussian" to the spatial neighborhood of the selected target Gaussian, and generates a clear three-dimensional Gaussian scene representation 4. A high-precision 3D Gaussian reconstruction system based on fuzzy perception and joint optimization, characterized by: The system includes: a module for executing a high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization according to any one of claims 1 to 3.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, a high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization as described in any one of claims 1 to 3 is implemented.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization as described in any one of claims 1 to 3 is implemented.

7. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the high-precision three-dimensional Gaussian reconstruction method based on fuzzy perception and joint optimization as described in any one of claims 1 to 3 is implemented.

Images