A physical specimen panoramic three-dimensional model automatic construction system and method

Abstract

The present application relates to the technical field of three-dimensional digital modeling, and particularly relates to a system and method for automatically constructing a panoramic three-dimensional model of a physical specimen, which realizes automatic and omnidirectional collection of specimen images and adaptive adjustment of camera magnification through a panoramic three-dimensional collection system, adopts a deep learning cutout algorithm to eliminate background interference and extract the foreground of the specimen, generates a sparse point cloud model through deep convolutional neural network feature extraction and pose optimization, constructs a high-precision Gaussian point cloud model through 3D Gaussian reconstruction, and finally completes high-fidelity real-time rendering of the model using Splatting technology; the present application realizes automatic construction of a specimen three-dimensional model in the whole process, greatly improves the modeling efficiency and precision, and the average modeling time consumption is only 13 minutes, and the average measurement error is as low as 0.03 mm, without the need for a large amount of manual operation and post-processing, and being suitable for specimen digitization requirements in the fields of earth science, biology and cultural relic protection.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional digital modeling technology, specifically to an automatic system and method for constructing panoramic three-dimensional models of physical specimens. Background Technology

[0002] Digital specimen 3D modeling technology is a key technology in fields such as earth science, biology, and cultural relic conservation. With the advent of the big data era, 3D digital models play a vital role in basic research and related applications due to their ability to provide rich and diverse data. This technology makes the digitization of physical specimens possible, providing new perspectives and methods for research and display in related fields. Currently, common physical specimen modeling techniques in this field mainly employ 3D scanners to digitize specimens and utilize specialized software for manual texturing and detail restoration.

[0003] Existing panoramic 3D model building technologies have many shortcomings in practical applications. First, the acquisition process requires a large amount of manual intervention, and the operator's skill level and subjective judgment directly affect the acquisition quality and efficiency. Moreover, manual operation is prone to errors, making it unsuitable for specimen acquisition tasks with high repetition and strict detail requirements. Second, the modeling quality is difficult to guarantee, as irrelevant objects or cluttered elements in the background environment may be incorrectly included in the model, affecting the model's reconstruction accuracy. In addition, after the model is built, a large amount of post-processing work is required, including manual removal and optimization of noise in the model, and the creation of surface textures to restore the model's true colors. The overall operation process is cumbersome, and the efficiency and quality of specimen 3D digitization cannot meet the industry's needs for large-scale and high-precision processing. Therefore, in view of the above situation, there is an urgent need to develop an automatic panoramic 3D model building system and method for physical specimens to overcome the shortcomings in current practical applications. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic system and method for constructing panoramic three-dimensional models of physical specimens, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] An automatic construction system for panoramic 3D models of physical specimens includes a panoramic 3D acquisition subsystem, an image processing unit, a 3D reconstruction unit, and a rendering unit. The image processing unit includes an image processing and foreground segmentation module, the 3D reconstruction unit includes a camera pose estimation and sparse reconstruction module and a 3D Gaussian reconstruction module, and the rendering unit includes a real-time rendering module.

[0007] The image processing and foreground segmentation module is communicatively connected to the panoramic 3D acquisition subsystem and is used to perform foreground segmentation on the specimen images acquired by the panoramic 3D acquisition subsystem, remove background interference, and obtain the specimen foreground image.

[0008] The camera pose estimation and sparse reconstruction module is communicatively connected to the image processing and foreground segmentation module, and is used to extract features and optimize camera pose of the specimen foreground image to generate a sparse point cloud model.

[0009] The 3D Gaussian reconstruction module is communicatively connected to the camera pose estimation and sparse reconstruction module, and is used to convert the sparse point cloud model into a 3D Gaussian model and construct a Gaussian point cloud model.

[0010] The real-time rendering module is communicatively connected to the 3D Gaussian reconstruction module and is used to render the Gaussian point cloud model in real time based on Splatting technology to obtain a three-dimensional digital specimen model.

[0011] As a further aspect of the present invention: the panoramic 3D acquisition subsystem includes a high-precision mechanical motion component, at least one industrial automatic zoom camera, and a core drive unit;

[0012] The core drive unit is equipped with a high-precision stepper motor, which drives the turntable to rotate step by step at a preset angle, and triggers the industrial automatic zoom camera to capture images synchronously when it rotates to the preset acquisition position.

[0013] As a further aspect of the present invention: the image processing and foreground segmentation module is equipped with a deep learning-based matting algorithm model, which includes an image encoder and a mask decoder;

[0014] The image encoder is used to perform feature encoding on the input specimen image, and the mask decoder is used to generate a segmentation mask based on the feature encoding, and extract the region where the specimen is located according to the segmentation mask to obtain a clean specimen foreground image.

[0015] As a further aspect of the present invention: the camera pose estimation and sparse reconstruction module includes a deep convolutional neural network feature extraction unit, a feature matching unit, and a pose optimization unit;

[0016] The deep convolutional neural network feature extraction unit is used to perform deep feature extraction and key point detection on the foreground image of the specimen.

[0017] The feature matching unit is used to match key point features between foreground images of specimens from different viewpoints and to remove mismatched features.

[0018] The pose optimization unit is used to calculate the initial pose of the camera based on the matched features, and optimize the camera pose using the bundle adjustment method to generate a sparse point cloud model.

[0019] As a further aspect of the present invention: the 3D Gaussian reconstruction module includes a Gaussian initialization unit, a gradient optimization unit, and an adaptive density control unit;

[0020] The Gaussian initialization unit is used to initialize each three-dimensional point in the sparse point cloud model as a Gaussian ellipsoid.

[0021] The gradient optimization unit is used to combine the camera pose and camera intrinsic parameters to calculate the loss between the restored image generated by the Gaussian ellipsoid projection and the real image, and to update the attribute parameters of the Gaussian ellipsoid based on the gradient descent method.

[0022] The adaptive density control unit is used to perform cloning, segmentation, or pruning of the Gaussian ellipsoid based on its gradient value and opacity to generate an optimized Gaussian point cloud model.

[0023] A method for automatically constructing a panoramic 3D model of a physical specimen includes the following steps:

[0024] S1: Perform omnidirectional image acquisition on the target specimen to obtain multi-view image data of the specimen;

[0025] S2: Perform foreground segmentation processing on the acquired multi-view image data, remove background interference, and obtain the foreground image of the specimen;

[0026] S3: Perform feature extraction and camera pose optimization on the foreground image of the specimen to generate a sparse point cloud model;

[0027] S4: Convert the sparse point cloud model into a 3D Gaussian model and construct a Gaussian point cloud model;

[0028] S5: The Gaussian point cloud model is rendered in real time based on Splatting technology to obtain a three-dimensional digital specimen model.

[0029] As a further aspect of the present invention: step S1 specifically includes:

[0030] S11: Collect the front and top views of the target specimen and extract the length, width and height dimensions of the target specimen;

[0031] S12: Dynamically adjust the camera's magnification according to the size parameters so that the proportion of the specimen in the camera's field of view meets a preset threshold.

[0032] S13: Control the turntable to rotate step by step at a preset angle. When it rotates to the designated acquisition position, it triggers the camera to acquire images synchronously. At the same time, it adaptively configures the depth-of-field fusion distance according to the current shooting magnification of the camera to obtain clear multi-view image data of the target specimen.

[0033] As a further aspect of the present invention: In step S2, a deep learning-based image matting algorithm is used to perform foreground segmentation processing, the specific process of which is as follows:

[0034] The input multi-view image data is feature-encoded by an image encoder, and a segmentation mask is output by a mask decoder based on the generated feature encoding.

[0035] After binarizing the segmentation mask, the region where the specimen is located is extracted by segmenting the maximum connected component to obtain the foreground image of the specimen.

[0036] As a further aspect of the present invention: step S3 specifically includes:

[0037] S31: Input the foreground image of the specimen into a pre-trained deep convolutional neural network to extract deep features of the image and identify key points in the image;

[0038] S32: Dynamically adjust the feature extraction strategy based on the number of identified key points, refine the coordinates of the selected key points and generate corresponding feature descriptors;

[0039] The dynamic adjustment strategy is as follows: when the number of detected key points is lower than the first preset threshold, the maximum number of feature points extracted is increased.

[0040] When the number of detected key points exceeds the second preset threshold, the feature point screening ratio threshold is reduced.

[0041] S33: Match feature descriptors between images from different viewpoints, remove mismatched features, and calculate the initial pose of the camera.

[0042] S34: Globally optimize the camera pose using the bundle adjustment method to generate a sparse point cloud model.

[0043] As a further aspect of the present invention: step S4 specifically includes:

[0044] S41: Initialize each 3D point in the sparse point cloud model as a Gaussian ellipsoid;

[0045] S42: Combine camera pose and camera intrinsic parameters to project the Gaussian ellipsoid into a two-dimensional restored image, calculate the loss value between the restored image and the real image, and update the attribute parameters of the Gaussian ellipsoid based on the gradient descent method.

[0046] S43: Based on the gradient value and opacity of the Gaussian ellipsoid, perform adaptive density control on the Gaussian ellipsoid by cloning, segmentation or pruning to generate an optimized Gaussian point cloud model.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] The panoramic 3D acquisition subsystem of this invention is equipped with high-precision mechanical motion components, multiple industrial automatic zoom cameras, and a core drive unit with high-precision stepper motors. It can realize all-round automated shooting of specimens. At the same time, through adaptive dynamic adjustment of camera magnification and adaptive configuration of depth-of-field synthesis function, it ensures high-quality imaging of specimens of different sizes, greatly reduces manual operation processes, avoids the impact of manual operation on acquisition quality, effectively improves acquisition efficiency and positioning accuracy, and is suitable for specimen acquisition tasks with high repeatability and strict detail requirements.

[0049] This invention utilizes machine vision technology and deep learning algorithms to achieve rapid analysis of acquired images and rapid construction of 3D digital specimen models. Simultaneously, through automated foreground segmentation, feature matching, model reconstruction, and rendering, it eliminates the need for extensive manual post-processing. Tests have verified that the average modeling time of this invention is only 13 minutes, compared to the average time of 3 hours and 24 minutes for existing structured light equipment and 16 hours for purely manual modeling, achieving an order-of-magnitude improvement in modeling efficiency.

[0050] This invention uses a deep learning-based image matting algorithm to accurately identify and segment the area where the specimen is located, effectively removing background clutter and preventing background elements from being incorrectly included in the model. Simultaneously, through a dynamically adjusted keypoint extraction strategy, it can adapt to the feature extraction requirements of specimens with weak textures, ensuring the accuracy of model reconstruction. Testing has verified that the average measurement error of this invention for a 5mm standard block is only 0.03mm, far superior to the 0.31mm average measurement error of existing structured light equipment, completely preserving the detailed information of the specimen.

[0051] This invention effectively avoids the risk of algorithm overfitting through optimized feature matching and screening strategies and 3D Gaussian adaptive density control algorithms, while improving the stability and reliability of the model under different conditions, ensuring that the construction system can obtain high-quality three-dimensional models in various environments and conditions.

[0052] This invention employs a graphics rendering scheme based on Splatting technology, which enables real-time rendering of 3D models with high resolution and high fidelity without the need for additional manual modeling and texture processing. The rendering effect is smooth and continuous, fully preserving the 3D morphology and detailed features of the digital specimen, and greatly improving the visual effect and application range of the model. Attached Figure Description

[0053] Figure 1This is an overall flowchart of the method for automatically constructing a panoramic three-dimensional model of a physical specimen in an embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram of the panoramic 3D acquisition system in an embodiment of the present invention.

[0055] Figure 3 This is a diagram showing the foreground extraction effect of a specimen in an embodiment of the present invention.

[0056] Figure 4 This is a diagram showing the camera spatial pose localization result in an embodiment of the present invention.

[0057] Figure 5 This is a sparse point cloud model diagram of a specimen in an embodiment of the present invention.

[0058] Figure 6 This is a 3D Gaussian rendering result of a specimen in an embodiment of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0061] Please see Figures 1-6 The present invention provides an automatic construction system and method for panoramic three-dimensional models of physical specimens. The system includes a panoramic three-dimensional acquisition subsystem, an image processing and foreground segmentation module, a camera pose estimation and sparse reconstruction module, a 3D Gaussian reconstruction module, and a real-time rendering module.

[0062] The structure of the panoramic 3D acquisition subsystem is as follows: Figure 2 As shown, the system is equipped with high-precision mechanical motion components, multiple industrial automatic zoom cameras, and a core drive unit. The core drive unit is equipped with a high-precision stepper motor, which is responsible for the rotation control of the turntable. This system can achieve omnidirectional imaging of the target specimen, reducing manual operation through automated acquisition and improving acquisition efficiency and positioning accuracy.

[0063] The following combination Figure 1 The flowchart shown illustrates the overall process of model construction, providing a detailed explanation of the automatic construction method for panoramic 3D models of physical specimens provided in this embodiment. The method includes the following steps:

[0064] Step 1: Use a panoramic 3D acquisition system to acquire images of the target specimen.

[0065] During the collection preparation phase, the system dynamically adjusts the camera magnification based on the actual size of the specimen: First, the system captures the specimen's front and top views using the camera and extracts the specimen's length (L), width (W), and height (H) dimensions. Then, based on these parameters, the system scales the camera to ensure that the specimen's proportion in the camera's field of view meets a preset minimum threshold. The camera magnification is calculated using the following formula:

[0066] Z (1);

[0067] In the formula: Z is the current camera magnification, L, W, and H are the length, width, and height of the sample, in mm, and I... w The field of view of the camera is 1 times the normal width, in mm.

[0068] During the acquisition phase, the turntable rotates stepwise according to a set angle, and triggers the camera to acquire images synchronously each time it rotates to a designated position. At the same time, the camera automatically starts the depth-of-field fusion function, and its depth-of-field fusion distance will be adaptively configured according to the current camera magnification to ensure that specimens of different sizes can achieve clear imaging.

[0069] This step uses adaptive dynamic adjustment of the camera magnification to adapt to the collection needs of specimens of different sizes, ensuring that the specimen always maintains an appropriate proportion in the camera's field of view, and avoiding problems such as incomplete imaging or insufficient resolution caused by differences in specimen size.

[0070] By using stepping rotation and synchronous image acquisition with the camera, it achieves full-range automated acquisition of specimen images and high-precision positioning of spatial coordinates. It eliminates the need for manual adjustment of shooting position and angle, greatly reduces manual operation, avoids the impact of operator skill level and subjective judgment on acquisition quality, and avoids acquisition errors caused by manual operation, thereby improving acquisition efficiency and positioning accuracy. It is suitable for specimen acquisition tasks with high repeatability and strict detail requirements.

[0071] By adaptively configuring the depth-of-field fusion distance, sufficient depth of field and sharpness are ensured for specimen imaging at different magnifications, providing high-quality raw data for subsequent image processing and model building, and ensuring the integrity and accuracy of image data.

[0072] Step 2: Use a deep learning-based image matting algorithm to remove background clutter and obtain the foreground image of the specimen.

[0073] The specimen foreground extraction result in this step is as follows: Figure 3As shown. The deep learning-based image matting algorithm model includes an image encoder and a mask decoder; the image encoder converts the input image into a feature code that the model can process; the mask decoder generates a segmentation mask based on the above feature code; then, by binarizing the segmentation mask and calculating the maximum connected component, the region where the specimen is located is segmented to obtain a clean specimen foreground image.

[0074] This step uses deep learning algorithms to automatically identify and separate the foreground specimen from the background environment. It can accurately identify and segment the area where the specimen is located, effectively remove interference from irrelevant objects or cluttered elements in the background, and avoid background elements being incorrectly included in the subsequent reconstructed model, thus greatly improving the modeling quality. At the same time, no manual background removal is required, reducing the workload of post-processing and further improving the efficiency of specimen 3D digitization.

[0075] Step 3: Extract features from the foreground image of the specimen to complete camera registration and sparse point cloud reconstruction.

[0076] The specimen foreground image obtained in step 2 is input into a pre-trained deep convolutional neural network for feature extraction. The camera spatial localization result in this step is as follows: Figure 4 As shown, the generated sparse point cloud model is as follows: Figure 5 As shown.

[0077] The specific process is as follows: The encoder of the pre-trained deep convolutional neural network is responsible for extracting deep features of the input image, and then the key point detection layer identifies key points in the image; then, non-maximum suppression is applied to filter out the most significant key points, and the extraction strategy is dynamically adjusted according to the number of key points after filtering: if the number of key points is less than 100, the maximum number of feature points extracted is automatically increased to ensure that sufficient effective feature points can be obtained even on the surface of weakly textured specimens; if the number of key points exceeds 1000, the filtering ratio threshold is reduced to reduce duplicate matching points and avoid the risk of algorithm overfitting.

[0078] After the key points are selected, their coordinates are refined and descriptors describing their local features are calculated. Then, the FLANN matcher is used to match the feature descriptors between different images, and the RANSAC algorithm is used to remove mismatches, thereby obtaining the initial position of the camera. Finally, the BA optimization algorithm is used to optimize the camera position and generate a sparse point cloud model.

[0079] This step uses a deep convolutional neural network to automate feature extraction and key point detection. Combined with a dynamically adjusted extraction strategy, it can adapt to specimens with different texture features. In particular, it can effectively obtain sufficient effective feature points for specimens with weak textures, while avoiding overfitting caused by too many feature points, thus improving the adaptability of feature matching and the robustness of model reconstruction.

[0080] By using the FLANN matcher to achieve efficient feature matching, combined with the RANSAC algorithm to eliminate mismatches, and then using the BA optimization algorithm to optimize the camera pose, high-precision positioning of the camera spatial position can be achieved, ensuring the accuracy of the generated sparse point cloud model, providing reliable basic data for subsequent high-precision 3D reconstruction, and ultimately improving the reconstruction accuracy of the specimen model while fully preserving the detailed information of the specimen.

[0081] Step 4: Convert the sparse point cloud model into a 3D Gaussian point cloud model to construct a high-precision Gaussian point cloud model.

[0082] The sparse point cloud model generated in step 3 is converted into a 3D Gaussian model. The specific process is as follows:

[0083] First, each 3D point is initialized as a Gaussian ellipsoid. Combining the camera's spatial position and well-known camera internal parameters (including focal length, pixel size, distortion parameters, etc.), the Gaussian ellipsoid is reconstructed into a 2D image, and the loss between the reconstructed image and the real image is calculated. Next, the properties of the Gaussian ellipsoid are updated using gradients, and the ellipsoid is cloned and pruned through adaptive density control. When the gradient value of a Gaussian point exceeds a set threshold of 0.002, it is segmented and cloned to improve local detail density. When the opacity of the Gaussian ellipsoid is lower than 0.01, it is directly removed to reduce model redundancy. Finally, a high-quality Gaussian point cloud model is generated.

[0084] This step uses 3D Gaussian reconstruction technology to transform sparse point clouds into Gaussian ellipsoid models. Combined with an adaptive density control strategy, local detail density can be increased in areas rich in specimen detail through cloning, fully preserving the specimen's microscopic morphology and detailed features. At the same time, redundant Gaussian ellipsoids are removed by pruning, simplifying the model size and improving subsequent rendering efficiency.

[0085] By optimizing the Gaussian ellipsoid properties through gradient descent, the reconstruction accuracy of the model can be significantly improved. At the same time, by optimizing the algorithm, the stability and reliability of the model under different conditions can be improved, enhancing the robustness of the specimen model reconstruction and ensuring that the construction system can obtain high-quality 3D models under various environments and conditions.

[0086] Step 5: Render the Gaussian point cloud model using splatting technology to obtain a real-time visualized 3D digital specimen model.

[0087] The 3D Gaussian rendering result of this step is as follows: Figure 6 As shown. The specific rendering process is as follows:

[0088] For each screen pixel, first calculate the color contribution and transparency of all Gaussian points affecting that pixel. The color of each point is weighted according to its influence on that pixel (usually related to distance and Gaussian kernel size). Then, use the alpha blending color synthesis formula to combine the contributions of all Gaussian points to obtain the pixel color. The color synthesis formula is shown below:

[0089]

[0090] In the formula: C is the composite pixel color, N is the number of Gaussian points that affect the pixel, α is the Gaussian point transparency, F is the Gaussian point color, and B is the background color.

[0091] The Splatting rendering technology used in this step makes the rendering effect smoother and more continuous, fully preserving the three-dimensional shape and details of the digital specimen. Without the need for additional manual modeling and texturing, the three-dimensional model can be rendered in real time with high resolution and high fidelity, which greatly reduces the manual cost of post-processing and significantly improves the visual effect and application range of the model.

[0092] To verify the beneficial effects of the technical solution of this embodiment, modeling tests were conducted on specimens of different sizes, and the results were compared with those of existing technical solutions. The test results are shown in the table below:

[0093] Table 1. Statistics on modeling time for different schemes

[0094] Structured light 3h 48min 3h 25min 3 hours 16 minutes 3h 03min 2 hours 40 minutes 3h 24min Artificial modeling 16h 16h 16h 16h 16h 16h This invention 13min 13min 14min 12min 14min 13min

[0095] As shown in Table 1 above, the average modeling time for structured light equipment is 3 hours and 24 minutes, the average modeling time for manual modeling is 16 hours, while the average modeling time of this invention is 13 minutes, thus improving modeling efficiency.

[0096] In addition, the accuracy was verified using a standard gauge block with a size of 5mm. By comparing the measured size of the gauge block with the modeling size, the average measurement error of the present invention was found to be 0.03mm, and the average error of the structured light device was 0.31mm. The present invention has achieved an improvement in modeling accuracy, as shown in Table 2.

[0097] Table 2. Statistical table of modeling accuracy for different schemes

[0098] Structured light / mm 5.45 5.23 5.33 5.24 5.28 5.35 5.31 This invention / mm 5.00 5.00 4.99 4.93 4.97 4.95 4.97

[0099] As can be seen from the above test data, the present invention achieves an order-of-magnitude improvement in modeling efficiency compared with existing structured light equipment and purely manual modeling, while also significantly improving modeling accuracy compared with existing structured light equipment, thus achieving a dual improvement in the efficiency and quality of specimen 3D digitization.

[0100] It should be noted that, in this invention, although the specification describes the embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An automatic panoramic 3D model construction system for physical specimens, comprising a panoramic 3D acquisition subsystem, an image processing unit, a 3D reconstruction unit, and a rendering unit, characterized in that, The image processing unit includes an image processing and foreground segmentation module; the 3D reconstruction unit includes a camera pose estimation and sparse reconstruction module and a 3D Gaussian reconstruction module; and the rendering unit includes a real-time rendering module. The image processing and foreground segmentation module is communicatively connected to the panoramic 3D acquisition subsystem and is used to perform foreground segmentation on the specimen images acquired by the panoramic 3D acquisition subsystem, remove background interference, and obtain the specimen foreground image. The camera pose estimation and sparse reconstruction module is communicatively connected to the image processing and foreground segmentation module, and is used to extract features and optimize camera pose of the specimen foreground image to generate a sparse point cloud model. The 3D Gaussian reconstruction module is communicatively connected to the camera pose estimation and sparse reconstruction module, and is used to convert the sparse point cloud model into a 3D Gaussian model and construct a Gaussian point cloud model. The real-time rendering module is communicatively connected to the 3D Gaussian reconstruction module and is used to render the Gaussian point cloud model in real time based on Splatting technology to obtain a three-dimensional digital specimen model.

2. The automatic construction system for panoramic three-dimensional models of physical specimens according to claim 1, characterized in that, The panoramic 3D acquisition subsystem includes a high-precision mechanical motion component, at least one industrial automatic zoom camera, and a core drive unit. The core drive unit is equipped with a high-precision stepper motor, which drives the turntable to rotate step by step at a preset angle, and triggers the industrial automatic zoom camera to capture images synchronously when it rotates to the preset acquisition position.

3. The automatic construction system for panoramic three-dimensional models of physical specimens according to claim 1, characterized in that, The image processing and foreground segmentation module is equipped with a deep learning-based matting algorithm model, which includes an image encoder and a mask decoder. The image encoder is used to perform feature encoding on the input specimen image, and the mask decoder is used to generate a segmentation mask based on the feature encoding, and extract the region where the specimen is located according to the segmentation mask to obtain a clean specimen foreground image.

4. The automatic construction system for panoramic three-dimensional models of physical specimens according to claim 1, characterized in that, The camera pose estimation and sparse reconstruction module includes a deep convolutional neural network feature extraction unit, a feature matching unit, and a pose optimization unit. The deep convolutional neural network feature extraction unit is used to perform deep feature extraction and key point detection on the foreground image of the specimen. The feature matching unit is used to match key point features between foreground images of specimens from different viewpoints and to remove mismatched features. The pose optimization unit is used to calculate the initial pose of the camera based on the matched features, and optimize the camera pose using the bundle adjustment method to generate a sparse point cloud model.

5. The automatic construction system for panoramic three-dimensional models of physical specimens according to claim 1, characterized in that, The 3D Gaussian reconstruction module includes a Gaussian initialization unit, a gradient optimization unit, and an adaptive density control unit; The Gaussian initialization unit is used to initialize each three-dimensional point in the sparse point cloud model as a Gaussian ellipsoid. The gradient optimization unit is used to combine the camera pose and camera intrinsic parameters to calculate the loss between the restored image generated by the Gaussian ellipsoid projection and the real image, and to update the attribute parameters of the Gaussian ellipsoid based on the gradient descent method. The adaptive density control unit is used to perform cloning, segmentation, or pruning of the Gaussian ellipsoid based on its gradient value and opacity to generate an optimized Gaussian point cloud model.

6. A method for automatically constructing a panoramic three-dimensional model of a physical specimen, characterized in that, Includes the following steps: S1: Perform omnidirectional image acquisition on the target specimen to obtain multi-view image data of the specimen; S2: Perform foreground segmentation processing on the acquired multi-view image data, remove background interference, and obtain the foreground image of the specimen; S3: Perform feature extraction and camera pose optimization on the foreground image of the specimen to generate a sparse point cloud model; S4: Convert the sparse point cloud model into a 3D Gaussian model and construct a Gaussian point cloud model; S5: The Gaussian point cloud model is rendered in real time based on Splatting technology to obtain a three-dimensional digital specimen model.

7. The method for automatically constructing a panoramic three-dimensional model of a physical specimen according to claim 6, characterized in that, Step S1 specifically includes: S11: Collect the front and top views of the target specimen and extract the length, width and height dimensions of the target specimen; S12: Dynamically adjust the camera's magnification according to the size parameters so that the proportion of the specimen in the camera's field of view meets a preset threshold. S13: Control the turntable to rotate step by step at a preset angle. When it rotates to the designated acquisition position, it triggers the camera to acquire images synchronously. At the same time, it adaptively configures the depth-of-field fusion distance according to the current shooting magnification of the camera to obtain clear multi-view image data of the target specimen.

8. The method for automatically constructing a panoramic three-dimensional model of a physical specimen according to claim 6, characterized in that, In step S2, a deep learning-based image matting algorithm is used to perform foreground segmentation processing. The specific process is as follows: The input multi-view image data is feature-encoded by an image encoder, and a segmentation mask is output by a mask decoder based on the generated feature encoding. After binarizing the segmentation mask, the region where the specimen is located is extracted by segmenting the maximum connected component to obtain the foreground image of the specimen.

9. The method for automatically constructing a panoramic three-dimensional model of a physical specimen according to claim 6, characterized in that, Step S3 specifically includes: S31: Input the foreground image of the specimen into a pre-trained deep convolutional neural network to extract deep features of the image and identify key points in the image; S32: Dynamically adjust the feature extraction strategy based on the number of identified key points, refine the coordinates of the selected key points and generate corresponding feature descriptors; The dynamic adjustment strategy is as follows: when the number of detected key points is lower than the first preset threshold, the maximum number of feature points extracted is increased. When the number of detected key points exceeds the second preset threshold, the feature point screening ratio threshold is reduced. S33: Match feature descriptors between images from different viewpoints, remove mismatched features, and calculate the initial pose of the camera. S34: Globally optimize the camera pose using the bundle adjustment method to generate a sparse point cloud model.

10. The method for automatically constructing a panoramic three-dimensional model of a physical specimen according to claim 6, characterized in that, Step S4 specifically includes: S41: Initialize each 3D point in the sparse point cloud model as a Gaussian ellipsoid; S42: Combine camera pose and camera intrinsic parameters to project the Gaussian ellipsoid into a two-dimensional restored image, calculate the loss value between the restored image and the real image, and update the attribute parameters of the Gaussian ellipsoid based on the gradient descent method. S43: Based on the gradient value and opacity of the Gaussian ellipsoid, perform adaptive density control on the Gaussian ellipsoid by cloning, segmentation or pruning to generate an optimized Gaussian point cloud model.

Images