A 3D reconstruction method of ancient artifacts based on single view and deep learning

Abstract

The present invention relates to the technical field of three-dimensional reconstruction of ancient cultural relics pictures, and discloses a method for three-dimensional reconstruction of ancient cultural relics based on single view and deep learning, comprising the following steps: Step 1: input a data set of ancient cultural relics and crafts similar to the ancient cultural relics; Step 2: according to the network structure of "encoder-decoder", use an encoder with a 3D-ResVGG network and a multi-path channel attention module to perform deep information mining and feature extraction on the data set to generate a three-dimensional network model of ancient cultural relics; Step 3: pre-process the ancient cultural relics pictures that need to be three-dimensionally reconstructed through an intelligent AI tool to generate a preset type of picture; Step 4: input the preset type of picture as a single view into the ancient cultural relics three-dimensional network model to generate a complete ancient cultural relics three-dimensional model of the ancient cultural relics pictures that need to be three-dimensionally reconstructed. The present invention proposes a new 3D-ResVGG network, which improves the problem of overfitting of the original VGG architecture through multi-branch model training.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional reconstruction of ancient cultural relic images, and in particular to a three-dimensional reconstruction method of ancient cultural relic images based on single view and deep learning. Background Art

[0002] Ancient cultural relics are relics and remains with historical, artistic and scientific value left by human beings in social activities. They are precious historical and cultural heritage of mankind. In many cases, people's understanding of ancient cultural relics is only through observation and understanding of pictures, which is not deep enough. Because we live in a three-dimensional world, but what the human eye sees is the two-dimensional projection of objects, and the form of objects obtained by devices such as cameras is usually a two-dimensional image. Therefore, three-dimensional reconstruction based on images is a core issue in many fields such as computer vision, computer animation, and industrial manufacturing. In the real world, all objects exist in three-dimensional form and have characteristics such as shape, appearance and texture. For engineering surveying and mapping technology, three-dimensional reconstruction can help its remote sensing system obtain the required surface space information and achieve real-time performance; in the field of human-computer interaction, three-dimensional reconstruction technology has laid the foundation for advanced human-computer interface and truly realized immersive and interactive virtual reality technology; in the field of building informatization, it is an important tool for cultural relics and building protection, engineering quality inspection and management, building demolition management, and building renovation or decoration.

[0003] At the same time, as a carrier for disseminating, preserving and displaying historical culture and social outlook to the public, the protection of cultural relics data has gradually been taken seriously by the public. Today, 3D reconstruction technology has gradually become the most advanced application technology in the preservation of historical cultural relics data. It can preserve the geometric information of real cultural relics for a long time, providing great convenience for promoting the application research and in-depth data analysis of cultural relics.

[0004] In addition, in the field of visualization technology, the digitization of physical cultural relics information and visualization technology in the computer field have been applied on a large scale. The relevant industries of culture and education will have more opportunities to give full play to the social education function of cultural relics to further expand and meet the public's appreciation needs.

[0005] Among the traditional methods of 3D reconstruction, the more mature one is the theory of stereo vision. Stereo vision is a scientific method based on the principle of pinhole imaging, which uses a camera to simulate the geometric law of human eyes viewing objects to calculate disparity and then deduce depth. In recent years, with the popularity of smart phones and the pursuit of lens blur by mobile phone manufacturers, 3D reconstruction based on binocular vision has become more and more mature. The disparity map is calculated by the dual camera of the camera, and the depth map is calculated by the disparity map to restore the 3D scene. At the same time, the addition of TOF sensors has also brought new solutions to this problem. TOF sensors can directly obtain sparse depth information. Although it is not accurate enough, it can be calculated by binocular vision. With the development of 3D reconstruction technology, a large number of stereo vision devices have been born, such as stereo cameras, Kinect, etc. However, despite the emergence of a large number of stereo vision devices, most of the ancient cultural relics images that are easily obtained in the network environment are still monocular single view, not RGBD containing depth or binocular view with given camera parameters. Single view images lacking related information are difficult to reconstruct through stereo vision theory, which also leads to a new computer vision problem: single view 3D reconstruction problem.

[0006] Single Vision Figure 3 The task of 3D reconstruction is a very difficult computer vision task. The traditional method of extracting features is mainly to construct manual features, use operators to extract image texture and contour, calculate optical flow to extract motion information, and judge the relationship between light and dark by grayscale. These are all common means of extracting manual features. Using these features plus auxiliary assumptions to complete the reconstruction of the 3D model, of course, limited by the amount of information contained in the features, the traditional method is not ideal. Summary of the invention

[0007] In view of the current scarcity of ancient artifact view images and 3D model datasets, a 3D reconstruction method for ancient artifacts based on single view and deep learning is proposed on the basis of existing research and the realization of ancient artifact 3D model datasets. After the 3D network model of ancient artifacts is trained, a model can be generated by inputting a single view image to achieve the mapping between 2D images and 3D models, so that 3D reconstruction can be performed in the case of a single view image. In addition, the algorithm does not require image annotations or classification labels for training, which overcomes the problems of lack of texture and wide baseline feature matching that could not be solved in the past.

[0008] The present invention is achieved through the following technical solutions:

[0009] A method for 3D reconstruction of ancient artifacts based on single view and deep learning, comprising the following steps:

[0010] Step 1: input a data set of ancient artifacts and crafts similar to the ancient artifacts, wherein the data set includes multi-angle multi-view images of the ancient artifacts and crafts similar to the ancient artifacts and a training voxel model;

[0011] Step 2: According to the “encoder-decoder” network structure, an encoder with a 3D-ResVGG network and a multi-path channel attention module is used to perform deep information mining and feature extraction on the dataset to generate a three-dimensional network model of ancient artifacts;

[0012] Step 3: Pre-process the images of ancient cultural relics that need to be 3D reconstructed through intelligent AI tools to generate preset type images;

[0013] Step 4: Input the preset type image as a single view into the ancient cultural relic three-dimensional network model to generate a complete ancient cultural relic three-dimensional model of the ancient cultural relic image that needs to be three-dimensionally reconstructed.

[0014] As an optimization, in step 2, the ancient cultural relic three-dimensional network model includes an encoder, an iterative convolution module and a decoder. The encoder is composed of a residual network structure combined with a multi-path channel attention module and a 3D-ResVGG network. The iterative convolution module is composed of a group of iterative convolution units, which are spatially distributed in a three-dimensional grid structure. Each of the iterative convolution units is responsible for reconstructing the final output voxel probability result.

[0015] As an optimization, the 3D-ResVGG network adds two groups of 1*1 convolutions to the forward propagation part of the ResNet network and is obtained by combining the ResNet residual module and the VGG network architecture.

[0016] As an optimization, the function of the multi-path channel attention module is: using the multi-path channel attention module, performing group convolution in the process of feature extraction in the 3D-ResVGG network, deeply mining the data information in the form of multi-path multi-group convolution kernels, and obtaining the deep information of the data.

[0017] As an optimization, the specific steps of the encoder in step 2 for deep information mining and feature extraction of the data set are:

[0018] Step 2.1, the 3D-ResVGG network receives a data set and performs feature extraction on the data set to generate a feature map;

[0019] Step 2.2, the multi-path channel attention module performs group extraction on the feature map generated by the 3D-ResVGG network: the feature map is divided into multiple groups according to the characteristics of the feature map and the convolution kernel is selected to perform group convolution on the feature map;

[0020] Step 2.3, perform a group of 1*1 convolutions on each group of feature maps after convolution, simulating the residual network structure, to obtain several groups of corresponding feature maps;

[0021] Step 2.4: Concatenate and combine several groups of feature maps obtained in step 2.3 to obtain a feature map with high scale, high dimension and richer information.

[0022] As an optimization, in step 2.2, 3 groups of information convolution kernels are selected to perform group convolution on the feature map.

[0023] As an optimization, the iterative convolution unit is composed of a long and short-time convolutional network, which selects a specific number of convolutional units according to the specific format of the input feature map, and iteratively convolves the feature map extracted by the encoder from a three-dimensional perspective. The voxel probability results of each convolution in the feature map are obtained in the form of a classification network and collected and retained as a set of three-dimensional probability information.

[0024] As an optimization, in step 2, the specific process of the iterative convolution unit performing convolution operations on the input multiple groups of feature maps is: the iterative convolution unit performs iterative convolution operations on the multiple groups of multi-angle feature maps to cyclically extract the information of the feature maps, dynamically updates the voxel probability involved in the iterative convolution unit, and finally combines the voxel probability data of all iterative convolution units to synthesize a set of three-dimensional voxel probability models.

[0025] As an optimization, in the iterative convolution module, the specific structure of the iterative convolution unit is determined by the resolution of the feature map, and the X*X*X iterative convolution unit is used to process data of the feature map with X resolution.

[0026] As an optimization, the reliability of the three-dimensional network model of ancient cultural relics is verified by a loss function, and the loss function is a feature regularization softmax function after the improvement of the cross entropy function.

[0027] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0028] 1. The present invention can generate a corresponding three-dimensional voxel model by inputting a single picture of an ancient cultural relic, complete the training operation of the three-dimensional network model of the ancient cultural relic through multiple two-dimensional images, and then use the three-dimensional network model of the ancient cultural relic to reconstruct the two-dimensional image to be tested.

[0029] 2. In order to solve the problems of low video memory occupancy and flexibility in traditional ResNet networks and the overfitting problem in traditional VGG architecture when the network number is deep, the present invention proposes a new 3D-ResVGG network. Based on the original ResNet and VGG network architectures, the network improves the 3*3 residual operator in ResNet into a group of 3*3 convolution combined with 1*1 convolution branches through multi-branch model training, and adds Relu after each group of large branches, thereby improving the overfitting problem of the original VGG architecture.

[0030] 3. The present invention improves the problems of traditional encoder networks, such as the inability to extract information deeply and the slow training speed, and improves the accuracy and depth of the overall three-dimensional reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0032] Figure 1 A schematic diagram of the structure of a three-dimensional model of an ancient cultural relic according to a three-dimensional reconstruction method of ancient cultural relics based on a single view and deep learning according to the present invention;

[0033] Figure 2 This is a flowchart of a method for three-dimensional reconstruction of ancient artifacts based on single view and deep learning according to the present invention;

[0034] Figure 3 This is a flowchart of an encoder in a method for 3D reconstruction of ancient artifacts based on single view and deep learning according to the present invention;

[0035] Figure 4 It is a structural diagram of the multi-path channel attention module in the present invention;

[0036] Figure 5 This is a diagram of the 3D-ResVGG network structure in the present invention. DETAILED DESCRIPTION

[0037] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0038] Example 1

[0039] A method for 3D reconstruction of ancient artifacts based on single view and deep learning, comprising the following steps:

[0040] Step 1: Input ancient artifacts and crafts similar to the ancient artifacts (how to determine whether they are similar?) Use the existing ShapeNetVox dataset, which includes a dataset of voxel models similar to ancient artifacts such as porcelain, lamps, bowls, tables and chairs. The dataset includes multi-angle and multi-view images of ancient artifacts and crafts similar to the ancient artifacts, as well as training voxel models; the dataset here can use the existing ShapeNetVox dataset, which includes voxel models similar to ancient artifacts such as porcelain, lamps, bowls, tables and chairs.

[0041] Step 2: According to the “encoder-decoder” network structure, an encoder with a 3D-ResVGG network and a multi-path channel attention module is used to perform deep information mining and feature extraction on the dataset to generate a three-dimensional network model of ancient artifacts;

[0042] Step 3: Pre-process the ancient artifact images that need to be 3D reconstructed using intelligent AI tools to generate preset type images; the intelligent AI tools here may include but are not limited to PhotoScissor, and use image processing tools such as PhotoScissor to perform segmentation, cropping and other pre-processing on the images.

[0043] Step 4: Input the preset type image as a single view into the ancient cultural relic three-dimensional network model to generate a complete ancient cultural relic three-dimensional model of the ancient cultural relic image that needs to be three-dimensionally reconstructed.

[0044] In this embodiment, in step 2, the ancient artifact three-dimensional network model includes an encoder, an iterative convolution module and a decoder. The encoder is composed of a residual network structure combined with a multi-path channel attention module and a 3D-ResVGG network. The multi-path channel attention module is an efficient multi-scale channel attention mechanism network. The input of the encoder is a two-dimensional image of a single view, and the output of the encoder is a two-dimensional feature vector, which needs to be converted into three-dimensional information;

[0045] The iterative convolution module uses a group of iterative convolution units, which are composed of a group of long and short-time convolution networks, that is, continuous long and short-time and LRU networks. The iterative convolution units are spatially distributed in a three-dimensional grid structure. Each of the iterative convolution units is responsible for reconstructing the voxel probability module of the final output. Specifically, the iterative convolution unit selects a specific number of convolution units according to the specific format of the network input image, for example, it is determined according to the spatial resolution of the image. If the spatial resolution is 32, there are 32*32*32 convolution units. The feature map extracted by the above method is iteratively convolved from a three-dimensional perspective, and the voxel probability result of each convolution in the feature map is obtained in the form of a classification network and collected and retained as a group of three-dimensional probability information. The specific format of the feature map is determined according to the size of the image, which is equivalent to cutting the image into many small blocks, and each small block represents the voxel probability of the location of the small block; the input of the decoder is the three-dimensional information processed by the iterative convolution module, and the output of the decoder is the three-dimensional predicted voxel occupancy of a single image.

[0046] like Figure 5 As shown, in this embodiment, the 3D-ResVGG network is obtained by adding two groups of 1*1 convolutions to the forward propagation part of the ResNet network and combining the ResNet residual module and the VGG network architecture. The 3D-ResVGG network combines the advantages of both the ResNet residual module and the VGG network architecture to achieve high-precision feature extraction based on only using 3*3 convolutions and 1*1 convolutions.

[0047] In this embodiment, the function of the multi-path channel attention module is: using the multi-path channel attention module, performing group convolution in the process of feature extraction in the 3D-ResVGG network, activating it through the activation function Relu after convolution, and deeply mining the data information in the form of multi-path multi-group convolution kernels to obtain the depth information of the data.

[0048] Specifically, Figure 3-4 As shown, in this embodiment, the specific steps of the encoder in step 2 for performing deep information mining and feature extraction on the data set are:

[0049] Step 2.1, the 3D-ResVGG network receives a data set and performs feature extraction on the data set to generate a feature map;

[0050] Step 2.2, the multi-path channel attention module performs group extraction on the feature map generated by the 3D-ResVGG network: according to the characteristics of the feature map, the feature map is divided into multiple groups according to the convolution kernel, and several suitable groups of convolution kernels are selected to perform group convolution on the feature map, so as to reduce the amount of parameters and improve the accuracy of feature map extraction; the characteristics of the feature map here specifically refer to the determination of dividing into three or four groups according to the number of feature maps, and the convolution kernel is selected according to the actual training requirements. For example, if the accuracy requirement is high, a lower number of convolution kernels can be selected, and if there is a requirement for time, a convolution kernel with a high number of kernels can be selected. For example, according to the characteristics of the feature map, the feature map is divided into multiple groups according to the convolution kernels 1*1, 3*3, 5*5, 7*7, and 9*9, and three suitable groups are selected for group convolution; the reduction of the parameter amount here refers to that in the three-dimensional neural network, if no grouping is performed, 12 feature maps are generated, and the 12 feature maps may be too large for the server to process at one time, but the grouping can be 3, 3, 3, and 3 groups, and the server can process 3 at a time.

[0051] Step 2.3: To prevent overfitting of the result, each group of the feature maps after convolution is subjected to a group of 1*1 convolutions simulating the residual network structure to obtain several corresponding groups of feature maps; for example, according to the specific example of step 2.2, three groups of feature maps can be obtained.

[0052] Step 2.4: Concatenate and combine several groups of feature maps after weighted and residual processing in steps 2.1-2.3 to obtain a feature map with high scale, high dimension and richer information.

[0053] Figure 4 In the figure, "input" is multiple sets of feature maps, the Split node groups multiple sets of feature maps, conv is the convolution kernel, path is the path, SE is the residual module, and "output" is the combined and spliced ​​feature map.

[0054] In this embodiment, in step 2, the specific process of the iterative convolution unit performing convolution operations on the multiple groups of input feature maps is as follows: the iterative convolution unit performs iterative convolution operations on the multiple groups of multi-angle feature maps to cyclically extract the information of the feature maps, and dynamically updates the voxel probabilities involved in the iterative convolution unit. In the case of multiple views, the processing of each view will update the currently generated voxel probability model to generate a more accurate probability model, and finally combine the voxel probability data of all iterative convolution units to synthesize a set of three-dimensional voxel probability models. It should be noted that the cyclic extraction of feature map information here refers to a hidden layer between the input layer and the output layer of the iterative convolution unit that re-processes the information of the generated feature map, wherein the hidden layer of the iterative convolution unit (long-short time network) is set in the three-dimensional grid. Among them, h represents the output of the hidden layer, i, j, k represent the three dimensions of the hidden layer output, R represents the set of real numbers, and N represents a small batch of n samples, each of which has h input features.

[0055] In this embodiment, in the iterative convolution module, the specific structure of the iterative convolution unit is determined by the resolution of the feature map, and the X*X*X iterative convolution unit is used to process data of the feature map with X resolution, where X is a positive integer.

[0056] In this embodiment, the reliability of the three-dimensional network model of ancient cultural relics is verified by a loss function, and the loss function is a feature regularization softmax function improved by the cross entropy function. On the basis of the existing A-softmax and L-softmax, data normalization processing is added. When the quality of multiple groups of data is different, feature extraction is performed according to the specific picture data situation to reduce problems such as sample attention failure caused by picture quality.

[0057] The specific formula is as follows:

[0058]

[0059] Among them, θ is defined as the interval angle, m is the introduced parameter factor, defined as the angle feature distance parameter, m is used to adjust the distance between features, fixed to 4, s is the scaling factor, fixed to 30, and n represents the total number of samples in a batch. During the training process, the model will process a batch of samples at a time and calculate the average loss of all samples in the batch; c means that the input of the function is the result of compounding c different linear functions, that is, c is the number of decision functions (linear functions), and c means that the input from the upper layer will pass through c decision functions, which means that the output values ​​of multiple nodes are accumulated and summed.

[0060] The encoder of the present invention is a 3D-ResVGG network improved on the existing ResNet network, combining the advantages of the ResNet residual module and the VGG network architecture, and realizing high-precision feature extraction based on only using 3*3 convolution and 1*1 convolution; the iterative convolution module is composed of a group of iterative convolution units, each of which generates a voxel probability, and a complete three-dimensional voxel probability model is formed after multi-view image input; the decoder is composed of a traditional resnet network, which is implemented using the prior art resnet. The traditional resnet network model can prevent the overfitting problem caused by too deep network learning, and also solves the problem of unstable voxel probability model accuracy that may be caused by the encoder and iterative convolution module, and finally the convolution neural network is formed. Among them, the input of the encoder is a two-dimensional image of a single view, and the output of the encoder is a two-dimensional feature vector, which needs to be converted into three-dimensional information; the iterative convolution module completes the voxel probability model distributed in the three-dimensional grid. The input of the decoder is the three-dimensional information processed by the iterative convolution module, and the output of the decoder is the three-dimensional predicted voxel occupancy of a single image.

[0061] The feasibility of the solution in Example 1 is verified by combining specific experiments, as described below:

[0062] 1) Experimental Dataset

[0063] ShapeNet dataset

[0064] This dataset consists of 3D CAD models of objects and is the largest 3D model dataset with rich annotations to date. It is organized under the WordNet classification and provides rich semantic annotations for each 3D model, including physical dimensions, keywords, etc. The annotations can be provided through a web-based interface to achieve data visualization of object attributes. ShapeNet contains more than 3 million models, of which 220,000 models are classified into 3,135 categories.

[0065] 2) Evaluation Criteria

[0066] IoU: Intersection-over-Union

[0067] The present invention uses IoU as the evaluation index, and the larger the evaluation index, the better. Specific formula: Area or Overlap represents the overlapping area, that is, the intersection area, and Area or Union represents the union area.

[0068] 3) Training process

[0069] The present invention uses ShapeNet data set to train the ancient cultural relic three-dimensional network model. The input is a 127*127*3 picture, and after passing through the ancient cultural relic three-dimensional network model, the output is a 32*32*32 voxel space.

[0070] In summary, the present invention example verifies the feasibility of the scheme in Example 1 through the above-mentioned experimental process, experimental data and experimental results. The three-dimensional object reconstruction algorithm provided by the present invention example has good three-dimensional reconstruction capability for two-dimensional images.

[0071] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for 3D reconstruction of ancient artifacts based on single view and deep learning, characterized in that: The steps include: Step 1: input a data set of ancient artifacts and crafts similar to the ancient artifacts, wherein the data set includes multi-angle multi-view images of the ancient artifacts and crafts similar to the ancient artifacts and a training voxel model; Step 2: According to the "encoder-decoder" network structure, an encoder with a 3D-ResVGG network and a multi-path channel attention module is used to perform deep information mining and feature extraction on the dataset to generate a three-dimensional network model of ancient artifacts; The 3D-ResVGG network is obtained by adding two groups of 1*1 convolutions to the forward propagation part of the ResNet network and combining the ResNet residual module and the VGG network architecture; Step 3: Pre-process the images of ancient cultural relics that need to be 3D reconstructed through intelligent AI tools to generate preset type images; Step 4: Input the preset type image as a single view into the ancient cultural relic three-dimensional network model to generate a complete ancient cultural relic three-dimensional model of the ancient cultural relic image that needs to be three-dimensionally reconstructed.

2. The method for 3D reconstruction of ancient artifacts based on single view and deep learning according to claim 1, characterized in that: In step 2, the ancient artifact three-dimensional network model includes an encoder, an iterative convolution module and a decoder. The encoder is composed of a residual network structure combined with a multi-path channel attention module and a 3D-ResVGG network. The iterative convolution module is composed of a group of iterative convolution units. The iterative convolution units are spatially distributed in a three-dimensional grid structure. Each of the iterative convolution units is responsible for reconstructing the final output voxel probability result.

3. The method for 3D reconstruction of ancient cultural relics based on single view and deep learning according to claim 2, characterized in that: The function of the multi-path channel attention module is to use the multi-path channel attention module to perform group convolution in the process of feature extraction in the 3D-ResVGG network, deeply mine the data information in the form of multi-path multi-group convolution kernels, and obtain the depth information of the data.

4. The method for 3D reconstruction of ancient artifacts based on single view and deep learning according to claim 3, characterized in that: The specific steps of the encoder in step 2 for deep information mining and feature extraction of the data set are: Step 2.1, the 3D-ResVGG network receives a data set and performs feature extraction on the data set to generate a feature map; Step 2.2, the multi-path channel attention module performs group extraction on the feature map generated by the 3D-ResVGG network: the feature map is divided into multiple groups according to the characteristics of the feature map and the convolution kernel is selected to perform group convolution on the feature map; Step 2.3, perform a group of 1*1 convolutions on each group of feature maps after convolution, simulating the residual network structure, to obtain several groups of corresponding feature maps; Step 2.4: Concatenate and combine several groups of feature maps obtained in step 2.3 to obtain a feature map with high scale, high dimension and richer information.

5. The method for 3D reconstruction of ancient artifacts based on single view and deep learning according to claim 4, characterized in that: In step 2.2, three groups of information convolution kernels are selected to perform group convolution on the feature map.

6. The method for 3D reconstruction of ancient artifacts based on single view and deep learning according to claim 5, characterized in that: The iterative convolution unit is composed of a long and short-time convolutional network. The long and short-time convolutional network selects a specific number of convolutional units according to the specific format of the input feature map, iteratively convolves the feature map extracted by the encoder from a three-dimensional perspective, obtains the voxel probability result of each convolution in the feature map in the form of a classification network, and collects and retains a set of three-dimensional probability information.

7. The method for 3D reconstruction of ancient artifacts based on single view and deep learning according to claim 6, characterized in that: In step 2, the specific process of the iterative convolution unit performing convolution operations on the input multiple groups of feature maps is as follows: the iterative convolution unit performs iterative convolution operations on the multiple groups of multi-angle feature maps to cyclically extract the information of the feature maps, dynamically updates the voxel probability involved in the iterative convolution unit, and finally combines the voxel probability data of all iterative convolution units to synthesize a set of three-dimensional voxel probability models.

8. The method for 3D reconstruction of ancient artifacts based on single view and deep learning according to claim 7, characterized in that: In the iterative convolution module, the specific structure of the iterative convolution unit is determined by the resolution of the feature map, and the X*X*X iterative convolution unit is used to process data of the feature map with X resolution.

9. A method for 3D reconstruction of ancient artifacts based on single view and deep learning according to any one of claims 1 to 8, characterized in that: The reliability of the three-dimensional network model of ancient cultural relics is verified by a loss function, and the loss function is a feature regularization softmax function after the improvement of the cross entropy function.

Images