Implicit function cultural relic 3D model reconstruction method based on deep residual shrinkage network
By constructing a deep residual shrinkage network and combining it with threshold branches to optimize the cultural relics 3D model reconstruction process, the problems of large noise influence and low reconstruction accuracy are solved, and fast and high-precision cultural relics 3D model reconstruction is achieved.
Patent Information
- Application Number
- CN202210965245.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-12
- Publication Date
- 2025-09-23
- Estimated Expiration
- 2042-08-12
AI Technical Summary
The existing technology has problems such as large noise influence, low reconstruction accuracy and slow speed when reconstructing three-dimensional models of cultural relics, especially poor reconstruction effect for human models with different body characteristics.
A method based on deep residual contraction network is adopted. By constructing a deep residual contraction network consisting of a residual contraction encoder and a convolutional decoder in series, combined with threshold branches to remove noise, the implicit function reconstruction process of the three-dimensional model of cultural relics is optimized, and the deep residual contraction network is used to train and generate high-precision three-dimensional models of cultural relics.
It has achieved the rapid and accurate reconstruction of the three-dimensional model of cultural relics based on the two-dimensional images of cultural relics, improved the surface fineness and contour integrity of the model, and increased the reconstruction speed and accuracy.
Smart Images
Figure CN115330944B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of computer graphics processing, and in particular relates to the reconstruction of three-dimensional models of cultural relics. Technical Background
[0002] Due to the age and difficulty of excavating cultural relics, the models of cultural relics excavated often have varying degrees of damage. In order to restore the cultural relics to their original appearance, the design of restoration plans for cultural relics has always been a hot topic in cultural relic protection.
[0003] By reconstructing three-dimensional models of cultural relics from two-dimensional images and generating three-dimensional models of cultural relics for human-computer interaction, it has greatly promoted people's understanding of ancient Chinese history and culture, and combined science and technology with culture to promote historical development.
[0004] At present, three-dimensional reconstruction has been widely used in fields such as house construction and games. Through parametric models and depth scanning equipment, relatively accurate three-dimensional models can be generated. However, compared with more regular object models, how to accurately reconstruct three-dimensional models of people with different body characteristics has become a research focus. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a method for reconstructing three-dimensional models of cultural relics based on an implicit function deep residual shrinkage network, which eliminates the influence of noise, has high reconstruction accuracy and fast speed.
[0006] The technical solution adopted to solve the above technical problems consists of the following steps:
[0007] (1) Select cultural relics data
[0008] Three groups of cultural relic data containing two-dimensional images of cultural relics and corresponding three-dimensional models of cultural relics were selected, including the front background-free image of the three-dimensional model of the cultural relic, the front mask image of the three-dimensional model of the cultural relic, the target file of the three-dimensional model of the cultural relic with a resolution of 30k, the general format file of the three-dimensional model of the cultural relic with a resolution of 30k containing the texture features, materials, and camera information of the cultural relic, the target file of the three-dimensional model of the cultural relic with a resolution of 100k, the general format file of the three-dimensional model of the cultural relic with a resolution of 100k containing the texture features, materials, and camera information of the cultural relic, the normal map of the three-dimensional model of the cultural relic with a resolution of 2k and its corresponding grayscale image, and the normal map of the three-dimensional model of the cultural relic with a resolution of 8k and its corresponding grayscale image.
[0009] (2) Rendering cultural relic data
[0010] Each set of cultural relic data is rendered to obtain 360 360° rendered images of the original cultural relic 3D model, 360 360° cultural relic model target files with texture features of the original cultural relic 3D model, 360 360° point set data files of the original cultural relic 3D model, and 360 normal vector maps of the original cultural relic 3D model under 360 lighting environments.
[0011] (3) Classification of cultural relics data
[0012] The rendered cultural relic data is divided into a training set and a test set. The training set is 75% of the cultural relic data, and the test set is 25% of the cultural relic data. There is no overlap between the training set and the test set data.
[0013] (4) Constructing a truth function
[0014] The truth function f*(X) is constructed for the three-dimensional model file of the cultural relics as follows:
[0015]
[0016] Where X is the coordinate of the point in the three-dimensional model space of the cultural relic.
[0017] (5) Constructing implicit functions
[0018] Construct the implicit function f(F(x),z(X)) as follows:
[0019] f(F(x),z(X))=P
[0020] x=π(X)
[0021]
[0022]
[0023] Where x is the coordinate of a point on the 2D image of the cultural relic, F(x) is the feature vector of point x on the 2D image of the cultural relic, z(X) is the distance between point X in the 3D model space of the cultural relic and the camera, P is the probability that point X in the 3D model space of the cultural relic is inside the 3D model of the cultural relic, and m is the number of 2D images in the input set of cultural relic images.
[0024] (6) Constructing a deep residual shrinkage network
[0025] The deep residual contraction network consists of a residual contraction encoder and a convolutional decoder in series.
[0026] (7) Training deep residual shrinkage network
[0027] The training set is input into the deep residual shrinkage network for training, and the loss function is determined as follows:
[0028]
[0029] Where i is the serial number of the point, n is the number of points in the three-dimensional model space of the cultural relic, and n is set to 10 6 ~10 8 , x i is the coordinate of the i-th point in the two-dimensional image of the cultural relic, X i is the coordinate of the i-th point in the three-dimensional model space of the cultural relic.
[0030] The Adam dynamic optimization method is used to adjust the deep residual shrinkage network. The learning rate of different training cycles is different during the training process. The learning rate is reduced to the current 0.1 after every 10 rounds of training. It is iterated until the loss function L converges to obtain a trained deep residual shrinkage network.
[0031] (8) Reconstruction of 3D models of cultural relics
[0032] The trained deep residual contraction network is tested using the test set. The front background-free image of the cultural relic 3D model and the front mask image of the cultural relic 3D model in the test set are input into the trained deep residual contraction network to obtain the reconstructed 3D model of the cultural relic.
[0033] In the step (6) of constructing a deep residual contraction network of the present invention, the residual contraction encoder is composed of a first residual contraction unit, a second residual contraction unit, a third residual contraction unit, and a fourth residual contraction unit connected in series in sequence.
[0034] The first residual contraction unit of the present invention is composed of a residual contraction branch and an identity mapping branch connected in parallel. The structures of the second residual contraction unit, the third residual contraction unit, and the fourth residual contraction unit are the same as the structure of the first residual contraction unit.
[0035] The residual shrinkage branch of the present invention is composed of a residual branch and a threshold branch connected in series, the residual branch is composed of a three-dimensional convolution layer a1 and a normalization layer b1, an activation function layer c1, a three-dimensional convolution layer a2, a normalization layer b2, and an activation function layer c2 connected in series in sequence, the threshold branch is composed of a fully connected layer d1 and a normalization layer b, an activation function layer c, a fully connected layer d2, and a squeeze function layer f connected in series in sequence, and the average value processing layer e is connected in parallel with the fully connected layer d1, the normalization layer b, the activation function layer c, the fully connected layer d2, and the squeeze function layer f connected in series in sequence.
[0036] The convolution kernel size of the three-dimensional convolution layer a1 and the three-dimensional convolution layer a2 of the present invention is 3×3.
[0037] The convolution channel of the first residual contraction unit of the present invention is 32 and the step size is 2, the convolution channel of the second residual contraction unit is 64 and the step size is 2, the convolution channel of the third residual contraction unit is 128 and the step size is 2, and the convolution channel of the fourth residual contraction unit is 256 and the step size is 1.
[0038] In the step (6) of constructing a deep residual shrinkage network of the present invention, the output y of the threshold branch is as follows:
[0039]
[0040]
[0041] Among them, k is the distance between the three-dimensional space point and the surface of the three-dimensional model of the cultural relic, τ is the threshold, α is the coefficient, α∈(0,1), N is the number of input features, which is a finite positive integer, k h,w,c It is the distance between the three-dimensional space point with width h, length w and channel number c and the surface of the three-dimensional model of the cultural relic.
[0042] The present invention performs masking on a two-dimensional image of a cultural relic without background to obtain a masked image, completing preprocessing of the two-dimensional cultural relic image, constructing a truth function based on the three-dimensional cultural relic model, constructing an implicit function representing the mapping relationship between the two-dimensional cultural relic image and the three-dimensional cultural relic model, and constructing a deep residual contraction network. By training the deep residual contraction network, the implicit function describing the three-dimensional cultural relic model is continuously optimized to obtain a trained deep residual network, and the three-dimensional cultural relic model is reconstructed based on the two-dimensional cultural relic image. The present invention adds a threshold branch to the deep residual contraction network, and removes non-obvious features by setting the threshold. This effectively eliminates the interference of noise generated during the training process on the experimental results, improves the fineness of the reconstructed three-dimensional cultural relic model surface and the integrity of the three-dimensional cultural relic model outline, and has the advantages of high reconstruction accuracy and high speed. The present invention can be applied to the technical field of reconstructing three-dimensional cultural relic models based on two-dimensional cultural relic images. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is a flow chart of an embodiment.
[0044] Figure 2 It is a structural diagram of the deep residual shrinkage network.
[0045] Figure 3 It is a structural diagram of the residual shrinkage unit of the deep residual shrinkage network.
[0046] Figure 4 It is a two-dimensional picture of cultural relics without background.
[0047] Figure 5 It is a front schematic diagram of the generated three-dimensional model of the cultural relic.
[0048] Figure 6 This is a schematic diagram of the back of the generated three-dimensional model of the cultural relic.
[0049] Figure 7 It is a side schematic diagram of the generated three-dimensional model of the cultural relic. DETAILED DESCRIPTION
[0050] The present invention will be further described in detail below with reference to the accompanying drawings and examples, but the present invention is not limited to the following embodiments.
[0051] Example 1
[0052] exist Figure 1 In this embodiment, the method for reconstructing a three-dimensional model of cultural relics based on an implicit function deep residual shrinkage network consists of the following steps:
[0053] (1) Select cultural relics data
[0054] Three groups of cultural relic data containing two-dimensional images of cultural relics and corresponding three-dimensional models of cultural relics were selected, including the front background-free image of the three-dimensional model of the cultural relic, the front mask image of the three-dimensional model of the cultural relic, the target file of the three-dimensional model of the cultural relic with a resolution of 30k, the general format file of the three-dimensional model of the cultural relic with a resolution of 30k containing texture features, materials, and camera information, the target file of the three-dimensional model of the cultural relic with a resolution of 100k, the general format file of the three-dimensional model of the cultural relic with a resolution of 100k containing texture features, materials, and camera information, the normal map of the three-dimensional model of the cultural relic with a resolution of 2k and its corresponding grayscale image, and the normal map of the three-dimensional model of the cultural relic with a resolution of 8k and its corresponding grayscale image.
[0055] (2) Rendering cultural relic data
[0056] Each set of cultural relic data is rendered to obtain 360 360° rendered images of the original cultural relic 3D model, 360 360° cultural relic model target files with texture features of the original cultural relic 3D model, 360 360° point set data files of the original cultural relic 3D model, and 360 normal vector maps of the original cultural relic 3D model under 360 lighting environments.
[0057] (3) Classification of cultural relics data
[0058] The rendered cultural relic data is divided into a training set and a test set. The training set is 75% of the cultural relic data, and the test set is 25% of the cultural relic data. There is no overlap between the training set and the test set data.
[0059] (4) Constructing a truth function
[0060] The truth function f*(X) is constructed for the three-dimensional model file of the cultural relics as follows:
[0061]
[0062] Where X is the coordinate of the point in the three-dimensional model space of the cultural relic.
[0063] (5) Constructing implicit functions
[0064] Construct the implicit function f(F(x),z(X)) as follows:
[0065] f(F(x),z(X))=P
[0066] x=π(X)
[0067]
[0068]
[0069] Where x is the coordinate of a point on the 2D image of the cultural relic, F(x) is the feature vector of point x on the 2D image of the cultural relic, z(x) is the distance between point x in the 3D model space of the cultural relic and the camera; P is the probability that point x in the 3D model space of the cultural relic is inside the 3D model of the cultural relic, and m is the number of 2D images of cultural relics in the input set.
[0070] (6) Constructing a deep residual shrinkage network
[0071] The deep residual contraction network is composed of a residual contraction encoder and a convolutional decoder in series. The convolutional decoder is composed of four deconvolution layers with the same structure connected in series.
[0072] exist Figure 2 In the embodiment, the residual contraction encoder is composed of a first residual contraction unit 1, a second residual contraction unit 2, a third residual contraction unit 3, and a fourth residual contraction unit 4 connected in series in sequence. The first residual contraction unit 1 is composed of a residual contraction branch and an identity mapping branch connected in parallel. The structures of the second residual contraction unit 2, the third residual contraction unit 3, and the fourth residual contraction unit 4 are the same as that of the first residual contraction unit 1. The convolution channel of the first residual contraction unit 1 is 32 and the stride is 2, the convolution channel of the second residual contraction unit 2 is 64 and the stride is 2, the convolution channel of the third residual contraction unit 3 is 128 and the stride is 2, and the convolution channel of the fourth residual contraction unit 4 is 256 and the stride is 1.
[0073] exist Figure 3 In this embodiment, the residual shrinkage branch consists of a residual branch and a threshold branch connected in series. The residual branch is composed of a 3D convolutional layer a1, a normalization layer b1, an activation function layer c1, a 3D convolutional layer a2, a normalization layer b2, and an activation function layer c2 connected in series. The convolution kernel size of the 3D convolutional layer a1 and the 3D convolutional layer a2 is 3×3. The threshold branch is composed of a fully connected layer d1, a normalization layer b, an activation function layer c, a fully connected layer d2, and a squeeze function layer f connected in series. The average processing layer e is connected in parallel with the fully connected layer d1, normalization layer b, activation function layer c, fully connected layer d2, and squeeze function layer f connected in series.
[0074] The output y of the threshold branch is as follows:
[0075]
[0076]
[0077] Where k is the distance between the three-dimensional space point and the surface of the three-dimensional model of the cultural relic, τ is the threshold, α is the coefficient, α∈(0,1), in this embodiment, α is 0.5, N is the number of input features, which is a finite positive integer, k h,w,c It is the distance between the three-dimensional space point with width h, length w and channel number c and the surface of the three-dimensional model of the cultural relic.
[0078] (7) Training deep residual shrinkage network
[0079] The training set is input into the deep residual shrinkage network for training, and the loss function is determined as follows:
[0080]
[0081] Where i is the serial number of the point, n is the number of points in the three-dimensional model space of the cultural relic, and n is set to 10 6 ~10 8 In this embodiment, the value of n is 10 7 , x i is the coordinate of the i-th point in the two-dimensional image of the cultural relic, X i is the coordinate of the i-th point in the three-dimensional model space of the cultural relic.
[0082] The Adam dynamic optimization method is used to adjust the deep residual shrinkage network. The learning rate of different training cycles is different during the training process. The learning rate is reduced to the current 0.1 after every 10 rounds of training. It is iterated until the loss function L converges to obtain a trained deep residual shrinkage network.
[0083] (8) Reconstruction of 3D models of cultural relics
[0084] The trained deep residual contraction network is tested using the test set. The front background-free image of the cultural relic 3D model and the front mask image of the cultural relic 3D model in the test set are input into the trained deep residual contraction network to obtain the reconstructed 3D model of the cultural relic.
[0085] Completed the implicit function cultural relic three-dimensional model reconstruction method based on deep residual shrinkage network.
[0086] Example 2
[0087] The implicit function cultural relic 3D model reconstruction method based on the deep residual shrinkage network of this embodiment consists of the following steps:
[0088] (1) Select cultural relics data
[0089] This step is the same as in Example 1.
[0090] (2) Rendering cultural relic data
[0091] This step is the same as in Example 1.
[0092] (3) Classification of cultural relics data
[0093] This step is the same as in Example 1.
[0094] (4) Constructing a truth function
[0095] This step is the same as in Example 1.
[0096] (5) Constructing implicit functions
[0097] This step is the same as in Example 1.
[0098] (6) Constructing a deep residual shrinkage network
[0099] The deep residual contraction network is composed of a residual contraction encoder and a convolutional decoder in series. The convolutional decoder is composed of four deconvolution layers with the same structure connected in series.
[0100] The residual contraction encoder of this embodiment is composed of a first residual contraction unit 1, a second residual contraction unit 2, a third residual contraction unit 3, and a fourth residual contraction unit 4 connected in series in sequence. The first residual contraction unit 1 is composed of a residual contraction branch and an identity mapping branch connected in parallel. The structures of the second residual contraction unit 2, the third residual contraction unit 3, and the fourth residual contraction unit 4 are the same as the structure of the first residual contraction unit 1. The convolution channel of the first residual contraction unit 1 is 32 and the stride is 2, the convolution channel of the second residual contraction unit 2 is 64 and the stride is 2, the convolution channel of the third residual contraction unit 3 is 128 and the stride is 2, and the convolution channel of the fourth residual contraction unit 4 is 256 and the stride is 1.
[0101] The residual shrinkage branch in this embodiment consists of a residual branch and a threshold branch connected in series. The residual branch is composed of a 3D convolutional layer a1, a normalization layer b1, an activation function layer c1, a 3D convolutional layer a2, a normalization layer b2, and an activation function layer c2 connected in series. The convolution kernel size of the 3D convolutional layer a1 and the 3D convolutional layer a2 is 3×3. The threshold branch is composed of a fully connected layer d1, a normalization layer b, an activation function layer c, a fully connected layer d2, and a squeeze function layer f connected in series. The average processing layer e is connected in parallel with the series of fully connected layers d1, normalization layer b, activation function layer c, fully connected layer d2, and squeeze function layer f.
[0102] The output y of the threshold branch is as follows:
[0103]
[0104]
[0105] Where k is the distance between the three-dimensional space point and the surface of the three-dimensional model of the cultural relic, τ is the threshold, α is the coefficient, α∈(0,1), in this embodiment, α is 0.1, N is the number of input features, which is a finite positive integer, k h,w,c It is the distance between the three-dimensional space point with width h, length w and channel number c and the surface of the three-dimensional model of the cultural relic.
[0106] (7) Training deep residual shrinkage network
[0107] The training set is input into the deep residual shrinkage network for training, and the loss function is determined as follows:
[0108]
[0109] Where i is the serial number of the point, n is the number of points in the three-dimensional model space of the cultural relics, and n is 10 6 ~10 8 In this embodiment, the value of n is 10 6 , x i is the coordinate of the i-th point in the two-dimensional image of the cultural relic, X i is the coordinate of the i-th point in the three-dimensional model space of the cultural relic.
[0110] The other steps of this step are the same as those in Example 1.
[0111] The other steps are the same as those in Example 1, and a reconstructed 3D model of the cultural relic is obtained. This completes the implicit function 3D model reconstruction method of cultural relic based on the deep residual shrinkage network.
[0112] Example 3
[0113] The implicit function cultural relic 3D model reconstruction method based on the deep residual shrinkage network of this embodiment consists of the following steps:
[0114] (1) Select cultural relics data
[0115] This step is the same as in Example 1.
[0116] (2) Rendering cultural relic data
[0117] This step is the same as in Example 1.
[0118] (3) Classification of cultural relics data
[0119] This step is the same as in Example 1.
[0120] (4) Constructing a truth function
[0121] This step is the same as in Example 1.
[0122] (5) Constructing implicit functions
[0123] This step is the same as in Example 1.
[0124] (6) Constructing a deep residual shrinkage network
[0125] The deep residual contraction network is composed of a residual contraction encoder and a convolutional decoder in series. The convolutional decoder is composed of four deconvolution layers with the same structure connected in series.
[0126] The residual contraction encoder of this embodiment is composed of a first residual contraction unit 1, a second residual contraction unit 2, a third residual contraction unit 3, and a fourth residual contraction unit 4 connected in series in sequence. The first residual contraction unit 1 is composed of a residual contraction branch and an identity mapping branch connected in parallel. The structures of the second residual contraction unit 2, the third residual contraction unit 3, and the fourth residual contraction unit 4 are the same as the structure of the first residual contraction unit 1. The convolution channel of the first residual contraction unit 1 is 32 and the stride is 2, the convolution channel of the second residual contraction unit 2 is 64 and the stride is 2, the convolution channel of the third residual contraction unit 3 is 128 and the stride is 2, and the convolution channel of the fourth residual contraction unit 4 is 256 and the stride is 1.
[0127] The residual shrinkage branch in this embodiment consists of a residual branch and a threshold branch connected in series. The residual branch is composed of a 3D convolutional layer a1, a normalization layer b1, an activation function layer c1, a 3D convolutional layer a2, a normalization layer b2, and an activation function layer c2 connected in series. The convolution kernel size of the 3D convolutional layer a1 and the 3D convolutional layer a2 is 3×3. The threshold branch is composed of a fully connected layer d1, a normalization layer b, an activation function layer c, a fully connected layer d2, and a squeeze function layer f connected in series. The average processing layer e is connected in parallel with the series of fully connected layers d1, normalization layer b, activation function layer c, fully connected layer d2, and squeeze function layer f.
[0128] The output y of the threshold branch is as follows:
[0129]
[0130]
[0131] Where k is the distance between the three-dimensional space point and the surface of the three-dimensional model of the cultural relic, τ is the threshold, α is the coefficient, α∈(0,1), in this embodiment, α is 0.9, N is the number of input features, which is a finite positive integer, k h,w,c It is the distance between the three-dimensional space point with width h, length w and channel number c and the surface of the three-dimensional model of the cultural relic.
[0132] (7) Training deep residual shrinkage network
[0133] The training set is input into the deep residual shrinkage network for training, and the loss function is determined as follows:
[0134]
[0135] Where i is the serial number of the point, n is the number of points in the three-dimensional model space of the cultural relics, and n is 10 6 ~10 8 In this embodiment, the value of n is 10 8 , x i is the coordinate of the i-th point in the two-dimensional image of the cultural relic, X i is the coordinate of the i-th point in the three-dimensional model space of the cultural relic.
[0136] The other steps of this step are the same as those in Example 1.
[0137] The other steps are the same as those in Example 1. A reconstructed three-dimensional model of the cultural relic is obtained, and the implicit function three-dimensional model reconstruction method of the cultural relic based on the deep residual shrinkage network is completed.
[0138] In order to verify the beneficial effects of the present invention, the inventors used a two-dimensional picture of cultural relics without background (see Figure 4 ) The method for reconstructing the three-dimensional model of cultural relics based on the implicit function of the deep residual shrinkage network of Example 1 of the present invention was used to carry out simulation experiments. The experimental results are shown in FIG. Figure 5 、 6 , 7. By Figure 5 It can be seen that the method of the present invention can improve the front details of the three-dimensional model of the cultural relic based on the two-dimensional picture of the cultural relic, and reconstruct the front of the three-dimensional model of the cultural relic with better morphological details; Figure 6 It can be seen that the method of the present invention can estimate the side shape of the three-dimensional model of the cultural relic based on the two-dimensional picture of the cultural relic and reconstruct the approximate side shape of the three-dimensional model of the cultural relic; Figure 7 It can be seen that the method of the present invention can predict the back shape of the three-dimensional model of the cultural relic based on the two-dimensional image of the cultural relic and reconstruct the complete back of the three-dimensional model of the cultural relic.
Claims
1. A method for reconstructing 3D models of cultural relics using an implicit function based on a deep residual shrinkage network, characterized in that It consists of the following steps: (1) Select cultural relics data Select 3 groups of cultural relic data containing 2D images of cultural relics and corresponding 3D models of cultural relics, including a front-side image of the 3D model of the cultural relic without background, a front-side mask image of the 3D model of the cultural relic, a 30k resolution target file of the cultural relic model, a 30k resolution general format file of the cultural relic model containing the texture features, material, and camera information of the cultural relic, a 100k resolution target file of the cultural relic model, a 100k resolution general format file of the cultural relic model containing the texture features, material, and camera information of the cultural relic, a 2k resolution normal map of the 3D model of the cultural relic and its corresponding grayscale image, and an 8k resolution normal map of the 3D model of the cultural relic and its corresponding grayscale image; (2) Rendering cultural relic data Render each set of cultural relic data to obtain 360 360-degree rendering images of the original cultural relic 3D model, 360 360-degree cultural relic model target files with texture features, 360 360-degree point set data files of the original cultural relic 3D model, and 360 normal vector maps of the original cultural relic 3D model under 360 different lighting environments. (3) Classification of cultural relics data The rendered cultural relic data is divided into a training set and a test set. The training set is 75% of the cultural relic data, and the test set is 25% of the cultural relic data. There is no overlap between the training set and the test set data. (4) Constructing a truth function The truth function f*(X) is constructed for the three-dimensional model file of the cultural relics as follows: Where X is the coordinate of the point in the three-dimensional model space of the cultural relic; (5) Constructing implicit functions Construct the implicit function f(F(x),z(X)) as follows: f(F(x),z(X))=P x=π(X) Where x is the coordinate of a point on the 2D image of the cultural relic, F(x) is the feature vector of point x on the 2D image of the cultural relic, z(X) is the distance between point X in the 3D model space of the cultural relic and the camera; P is the probability that point X in the 3D model space of the cultural relic is inside the 3D model of the cultural relic, and m is the number of 2D images of the cultural relic in the input set. (6) Constructing a deep residual shrinkage network The deep residual contraction network consists of a residual contraction encoder and a convolutional decoder in series; (7) Training deep residual shrinkage network The training set is input into the deep residual shrinkage network for training, and the loss function is determined as follows: Where i is the serial number of the point, n is the number of points in the three-dimensional model space of the cultural relic, and n is set to 10 6 ~10 8 , x i is the coordinate of the i-th point in the two-dimensional image of the cultural relic, X i is the coordinate of the i-th point in the three-dimensional model space of the cultural relic; The Adam dynamic optimization method is used to adjust the deep residual shrinkage network. The learning rate of different training cycles is different during the training process. The learning rate is reduced to the current 0.1 after every 10 training rounds. It is iterated until the loss function L converges to obtain a trained deep residual shrinkage network. (8) Reconstruction of 3D models of cultural relics The trained deep residual contraction network is tested using the test set. The front background-free image of the cultural relic 3D model and the front mask image of the cultural relic 3D model in the test set are input into the trained deep residual contraction network to obtain the reconstructed 3D model of the cultural relic.
2. The method for reconstructing a 3D model of cultural relics using an implicit function based on a deep residual shrinkage network according to claim 1, wherein: In the step (6) of constructing a deep residual contraction network, the residual contraction encoder is composed of a first residual contraction unit (1), a second residual contraction unit (2), a third residual contraction unit (3), and a fourth residual contraction unit (4) connected in series in sequence.
3. The method for reconstructing a 3D model of cultural relics using an implicit function based on a deep residual shrinkage network according to claim 2, wherein: The first residual contraction unit (1) is composed of a residual contraction branch and an identity mapping branch connected in parallel; the structures of the second residual contraction unit (2), the third residual contraction unit (3), and the fourth residual contraction unit (4) are the same as the structure of the first residual contraction unit (1).
4. The method for reconstructing 3D models of cultural relics using an implicit function based on a deep residual shrinkage network according to claim 3, characterized in that: The residual shrinkage branch is composed of a residual branch and a threshold branch connected in series. The residual branch is composed of a three-dimensional convolution layer a1 and a normalization layer b1, an activation function layer c1, a three-dimensional convolution layer a2, a normalization layer b2, and an activation function layer c2 connected in series in sequence. The threshold branch is composed of a fully connected layer d1 and a normalization layer b, an activation function layer c, a fully connected layer d2, and a squeeze function layer f connected in series in sequence. The average value processing layer e is connected in parallel with the fully connected layer d1, the normalization layer b, the activation function layer c, the fully connected layer d2, and the squeeze function layer f connected in series in sequence.
5. The method for reconstructing a 3D model of cultural relics using an implicit function based on a deep residual shrinkage network according to claim 4, characterized in that: The convolution kernel size of the three-dimensional convolution layer a1 and the three-dimensional convolution layer a2 is 3×3.
6. The method for reconstructing a 3D model of cultural relics using an implicit function based on a deep residual shrinkage network according to claim 2 or 3, characterized in that: The convolution channel of the first residual contraction unit (1) is 32 and the step size is 2, the convolution channel of the second residual contraction unit (2) is 64 and the step size is 2, the convolution channel of the third residual contraction unit (3) is 128 and the step size is 2, and the convolution channel of the fourth residual contraction unit (4) is 256 and the step size is 1.
7. The method for reconstructing a 3D model of cultural relics based on an implicit function deep residual shrinkage network according to claim 4, characterized in that The output y of the threshold branch is as follows: Among them, k is the distance between the three-dimensional space point and the surface of the three-dimensional model of the cultural relic, τ is the threshold, α is the coefficient, α∈(0,1), N is the number of input features, which is a finite positive integer, k h,w,c It is the distance between the three-dimensional space point with width h, length w and channel number c and the surface of the three-dimensional model of the cultural relic.